Creation contribution degree evaluation system and method based on keyboard behavior and AI interaction content

Abstract

The invention discloses a creation contribution degree evaluation system and method based on keyboard behaviors and AI interaction content, and relates to the technical field of computer software and information security. The method comprises the following steps: collecting keyboard input rhythm, rollback behavior and paste operation data in a user creation process; obtaining interactive dialogue content between the user and the AI model; generating an independent authoring confidence rating based on behavioral data analysis; performing semantic comparison on the user creation text and the AI interaction content, and calculating a human semantic contribution degree; fusing the confidence rating and the semantic contribution degree to generate a human creation contribution degree index; and finally, carrying out block chain evidence storage on the contribution statement abstract. According to the method, through a double evidence chain of'behavior + semantics', accurate quantification of the contribution ratio of human and AI is realized from the creation process, the problems of easy avoidance, high misjudgment rate and lack of credible evidence storage in the prior art are effectively solved, and the method is suitable for evaluation requirements of the fields of scientific research, education, publishing and the like on creation authenticity.

Description

Technical Field

[0001] This invention relates to the fields of computer software and information security technology, and in particular to a system and method for evaluating the creative contribution of content based on keyboard behavior and AI interaction. Background Technology

[0002] Current AI-assisted creative writing detection technologies primarily rely on probabilistic judgments of the final text. This involves analyzing static features such as grammatical structure, word frequency, and semantic patterns using pre-trained models, comparing these features with a known AI-generated text feature library, and outputting an "AI generation probability" value. However, existing technologies have significant drawbacks: their detection logic is simplistic and easily circumvented; users can significantly alter text features through simple synonym replacement, paragraph restructuring, or minor editing, thus reducing the AI ​​recognition rate. Furthermore, the technology suffers from a high false positive rate, struggling to accurately distinguish between scenarios like "human imitation of AI-style writing" and "deep rewriting after AI generation." Additionally, due to its reliance on a fixed feature library, its update speed cannot keep pace with the rapid iteration of AI models, resulting in insufficient detection capabilities for newly generated text. Most importantly, existing technologies only provide vague probability values, failing to quantify the specific contributions of humans and AI in the creative process, and lacking an immutable and reliable evidence preservation mechanism. This fails to meet the stringent requirements for accurate verification of creative authenticity in fields such as research, teaching, and publishing. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a creative contribution evaluation system and method that can accurately quantify and reliably verify the contributions of AI and humans in collaborative creation from the perspective of the creative process and integrate both behavioral and content dimensions. To achieve the above objectives, the present invention adopts the following technical solution: A method for evaluating the contribution of content creation based on keyboard behavior and AI interaction includes the following steps: Collect data on keyboard input rhythm, backtracking behavior, and paste operation during the user's creation process; Obtain the interactive dialogue content between the user and the AI ​​model; An independent creative confidence rating is generated based on the analysis of collected behavioral data. The semantics of user-generated text are compared with the acquired AI interactive content to calculate the human semantic contribution. By integrating the independent creative confidence rating with the human semantic contribution, a human creative contribution index is generated; The summary of contribution statements, which includes an index of human creative contribution, will be stored on the blockchain.

[0004] Preferably, the acquisition of keyboard input rhythm includes: A sliding window mechanism is used to record the time interval of continuous keyboard input; The standard deviation of the input rhythm and the proportion of long pauses are calculated based on time intervals. Based on a preset rhythmic characteristic baseline, determine whether the input rhythm is a human-contributed operation or a mechanical operation.

[0005] Preferably, the data rollback behavior includes: Distinguish between deleting keyboard input and deleting pasted content; The effective rollback rate is calculated based on the deletion of keyboard input. Analyze the relationship between fallback behavior and document structure to identify structural hierarchy fallback and semantic unit fallback.

[0006] Preferably, the collection and pasting operation includes: The paste operation is determined by pre-caching the clipboard content and comparing it with the text. Distinguish between AI-generated and non-AI-generated paste; Calculate the post-paste modification rate, which is the proportion of pasted content that is subsequently deleted.

[0007] Preferably, the acquisition of the interactive dialogue content between the user and the AI ​​model includes: Extract conversation content in real time from a designated AI service webpage using a browser extension; Supports importing official exported dialogue files from AI platforms; OCR recognition is used to extract the dialogue text from screenshots of conversations. Assign credibility weights to dialogue content from different sources.

[0008] Preferably, the generation of an independent creative confidence rating includes: The rhythm anomaly and regression anomaly are calculated based on the preset average behavior baseline; The pasting mode is determined based on the AI ​​content matching rate and the post-paste modification rate, including AI direct reuse mode, AI critical modification mode and non-AI pasting mode. Based on the anomaly calculation results and the paste mode judgment results, a three-level independent creation confidence rating of high, medium and low is output.

[0009] Preferably, the calculation of human semantic contribution includes: Intelligent text segmentation based on semantic boundaries; Convert text fragments into vector representations and calculate similarity; Set a similarity threshold to distinguish between content with high similarity, medium similarity, and low similarity; Based on the similarity calculation results, the human semantic contribution is calculated from three dimensions: original contribution, rewriting contribution, and structural reorganization contribution.

[0010] Preferably, the generation of the human creative contribution index includes: Transform independent creative confidence ratings into behavioral scores; The fusion weights of the behavioral and semantic dimensions are dynamically assigned based on confidence ratings and semantic analysis confidence levels. Cross-validation correction is performed based on the correlation features between behavior and semantics; The human creative contribution index, ranging from 0.0 to 100.0, is obtained through weighted fusion and correction.

[0011] This invention also proposes a system for evaluating the creative contribution of content based on keyboard behavior and AI interaction, comprising: The local editing behavior collection module is configured to collect data on keyboard input rhythm, backtracking rate, and paste block distribution during the user's creation process. The AI ​​dialogue interception and caching module is configured to capture or import user interaction content with AI in real time. The behavioral fingerprint generation and independent thinking index calculation module is configured to generate an independent thinking index based on behavioral characteristics. The AI ​​text semantic comparison and contribution calculation module is configured to compare user text with AI interaction content semantically and calculate semantic contribution. The dual-engine fusion module is configured to fuse behavioral indices and semantic contributions to generate a human creative contribution index. The blockchain evidence storage module is configured to upload contribution statement summaries to the blockchain.

[0012] Preferably, The local editing behavior acquisition module includes a keyboard input rhythm acquisition unit, a rollback rate calculation unit, and a paste block distribution recognition unit; The AI ​​dialogue interception and caching module includes a browser extension unit, a file import unit, and a confidence level unit; The behavioral fingerprint generation and independent thinking index calculation module includes an anomaly calculation unit, a pasting pattern diagnosis unit, and a confidence rating unit. The AI ​​text semantic comparison and contribution calculation module includes a text segmentation unit, a sentence vector matching unit, and a contribution hierarchical calculation unit. The dual-engine fusion module includes a dynamic weight calculation unit, a cross-validation unit, and a human creative contribution index calculation unit. The blockchain evidence storage module supports uploading the hash value of the contribution statement summary to the consortium blockchain or public test chain.

[0013] The beneficial effects of this invention are: This invention upgrades the process from "probabilistic guessing of results" to "credible quantification of the process" by collecting keyboard behavior features during the creation process and intercepting AI interaction content. It achieves, for the first time, a traceable and quantifiable assessment of the intensity and type of human thought. It constructs a dual evidence chain of "behavior + semantics," effectively preventing users from circumventing detection through superficial modifications. A dual-engine cross-validation mechanism significantly reduces the misjudgment rate for legitimate scenarios such as "humans deeply rewriting AI content." Simultaneously, based on blockchain-based zero-knowledge storage technology, it achieves verifiability of evaluation results without disclosing the original content, resolving the contradiction between privacy protection and third-party verification. This technical solution is based on the commonalities of human creative behavior and human-computer interaction, rather than relying on a specific AI text feature library, thus possessing stronger adaptability to technological evolution and long-term effectiveness. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the creation contribution evaluation method based on keyboard behavior and AI interaction content proposed in this invention.

[0016] Figure 2 This is a block diagram of the creation contribution evaluation system based on keyboard behavior and AI interaction content proposed in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0018] The creation contribution evaluation system and method based on keyboard behavior and AI interaction content proposed in this invention are executed by computer equipment.

[0019] Reference Figure 1The flowchart of the content creation contribution evaluation method based on keyboard behavior and AI interaction proposed in this invention is shown, including the following steps: S1: Collect data on keyboard input rhythm, backtracking behavior, and paste operation during the user's creation process; Specifically, the acquisition of keyboard input rhythm includes: S11: A sliding window mechanism is used to record the time interval of continuous keyboard input. User authorization is completed upon system startup, and a local authoring environment based on a text editor is initiated. A thread-safe circular queue is used as the sliding window, with a fixed queue length of 400 records. Each valid keyboard input (i.e., a character increment not triggered by undo / redo or paste) is recorded, the current timestamp is calculated, and the time interval between the current input and the previous input is calculated and stored in the queue. When the queue is full, the oldest interval is removed, achieving a sliding update of the data. S12: Calculate the standard deviation of the input rhythm and the proportion of long pauses based on the time interval; the background thread periodically counts all time intervals within the current sliding window, calculates their standard deviation to reflect the fluctuation of the input rhythm, and counts the proportion of long pauses greater than or equal to a preset threshold (such as 1500 milliseconds) to the total number of times. S13: Based on the preset rhythmic characteristic baseline, determine whether the input rhythm belongs to human contribution operation or mechanical operation; compare the calculated standard deviation and long pause ratio with the preset baseline. For example, the preset baseline may be set as follows: if the standard deviation is greater than or equal to 500 milliseconds and the long pause ratio is greater than or equal to 20%, it tends to be judged as a human contribution operation accompanied by thinking; if the standard deviation is less than or equal to 100 milliseconds and the long pause ratio is less than or equal to 5%, it tends to be judged as a mechanical operation or direct copying.

[0020] Specifically, the data rollback behavior includes: S14: Differentiate between deleting keyboard input and deleting pasted content; the system maintains a global event stream containing all authoring operations (input, paste, delete) and marks the content source of each event. When a deletion operation occurs, based on the source mark of the deleted text, it is distinguished as deleting content entered by the user's own keyboard (Category A deletion) or deleting content from block operations such as paste (Category B deletion); S15: Calculate the effective rollback rate based on the deletion of keyboard input; within the sliding window, accumulate the total number of characters deleted in category A, and divide it by the total number of characters added via the keyboard within the same window to obtain the effective rollback rate. This metric reflects the intensity of user modifications during their own input process.

[0021] S16: Analyze the relationship between rollback behavior and document structure, identifying structural-level rollback and semantic-unit rollback; by analyzing the context in which the cursor is located when deletion occurs (e.g., whether it is at the beginning of a heading, list item, or paragraph), determine whether the deletion involves adjustments to the document's logical structure (structural-level rollback) or only modifications to the wording within a sentence (semantic-unit rollback). Simultaneously, the creation process of a text unit can be divided into early, middle, and late stages according to a timeline, and the percentage of rollback characters in each stage can be statistically analyzed to identify the user's rollback pattern.

[0022] Specifically, the data collection and pasting operation includes: S17: Paste operations are determined by pre-caching clipboard content and comparing the text; when the editor gains focus or a paste shortcut is detected, the text content of the system clipboard is immediately read, a hash value snapshot is generated and cached. When the editor content changes significantly, the newly added text is compared with the cached clipboard snapshot; if the text overlaps significantly, it is determined to be a paste operation. S18: Distinguish between AI-sourced and non-AI-sourced pastes; the system will perform semantic similarity matching between the determined paste content and the AI ​​interaction content cache obtained through step S2. If a high-confidence AI-generated fragment is matched, the paste is marked as AI-sourced; otherwise, it is marked as non-AI-sourced (possibly from references, previous personal works, etc.). S19: Calculate the post-paste modification rate, which is the proportion of pasted content that was subsequently deleted. For each paste operation, the system tracks the number of characters deleted from the pasted content area during subsequent editing and calculates its percentage relative to the original pasted characters. A high modification rate indicates that the user has deeply critiqued and rewritten the externally imported content.

[0023] S2: Obtain the interactive dialogue content between the user and the AI ​​model; specifically, obtaining the interactive dialogue content between the user and the AI ​​model includes: S21: Extract dialogue content in real time from designated AI service web pages via browser extensions; develop lightweight browser extensions that, when users visit preset AI service web pages (such as ChatGPT), automatically identify and extract text pairs of "user questions" and "AI replies" by listening to changes in the page DOM, send them to the local evaluation system through a secure local communication channel, and mark them as high-confidence sources. S22: Import official exported dialogue files from AI platforms; provides a file import interface to support parsing official exported dialogue files (such as JSON format) provided by mainstream AI platforms such as ChatGPT, automatically extracts structured dialogue history, and marks them as sources with medium confidence. S23: Perform OCR recognition on the dialogue screenshot to extract the dialogue text; for AI dialogue screenshots provided by users, use the local OCR engine to recognize the text in the image, and then use a lightweight model to help distinguish between user statements and AI replies, extract the dialogue content, and mark it as a low-confidence source; S24: Assign confidence weights to dialogue content from different sources; For dialogue data obtained from the above different channels, assign different confidence weights (e.g., browser extension records are 1.0, official exported files are 0.6, and OCR recognition results are 0.3) based on the reliability and completeness of their sources, for weighted calculation in subsequent analysis.

[0024] S3: Generate an independent creative confidence rating based on the collected behavioral data analysis; specifically, the generation of the independent creative confidence rating includes: S31: Calculate rhythm anomaly and regression anomaly based on the preset average behavior baseline; the system adopts a fixed average baseline derived from a large number of sample statistics (e.g., the average input rhythm standard deviation is 150 milliseconds and the average regression rate is 15%); divide the current user's actual rhythm standard deviation and regression rate by the corresponding average baseline value to obtain the rhythm anomaly and regression anomaly, which are used to quantify the degree to which their behavior deviates from the average level of ordinary creators; S32: Determine the pasting mode based on the AI ​​content matching rate and the post-paste modification rate; combine the results of S18 and S19 to determine the user's pasting behavior mode: if the proportion of AI-source pasting is high and the modification rate is low, it is determined to be the "AI direct reuse" mode; if the proportion of AI-source pasting is moderate but the modification rate is high, it is determined to be the "AI critical modification" mode; if the proportion of AI-source pasting is very low, it is determined to be the "non-AI pasting" mode. S33: Combining the anomaly calculation results and the paste mode judgment results, output a three-level independent creation confidence rating of high, medium, and low; perform a comprehensive rating according to preset rules: for example, if the paste mode is "AI direct reuse" or the behavior anomaly is significantly low (indicating mechanical behavior), then output a "low" confidence rating; if the paste mode is "AI critical modification" and the behavior anomaly is significantly high (indicating high thinking intensity), then output a "high" confidence rating; otherwise, output a "medium" confidence rating.

[0025] S4: Semantically compare the user-generated text with the acquired AI interactive content to calculate the human semantic contribution; specifically, the calculation of the human semantic contribution includes: S41: Intelligent text segmentation based on semantic boundaries; according to language characteristics, the final authored text submitted by the user is segmented first according to semantic boundaries such as sentences and paragraphs. A maximum length limit is also set to prevent excessively long individual semantic segments from affecting analysis efficiency. S42: Convert text fragments into vector representations and calculate similarity; using a lightweight sentence vector model, convert each fragment of the user text and the AI ​​response fragments in the AI ​​dialogue content obtained in S2 into high-dimensional vectors respectively, and then calculate the cosine similarity between each fragment of the user text and all AI response fragments.

[0026] S43: Set similarity thresholds to distinguish between high, medium, and low similarity content; set two similarity thresholds (e.g., 0.90 and 0.70). A similarity of 0.90 or higher is considered "high similarity," indicating that the AI ​​content was almost entirely preserved; a similarity between 0.70 and 0.90 is considered "medium similarity," indicating that the AI ​​content was partially rewritten; a similarity below 0.70 is considered "low similarity," indicating that the content has high originality or has been deeply rewritten. S44: Based on the similarity calculation results, calculate the human semantic contribution from three dimensions: original contribution, rewriting contribution, and structural recombination contribution; Original contribution: The proportion of characters in "low similarity" text segments to the total text is calculated and mapped to an original contribution score; Rewriting contribution: For text segments with "medium similarity", the rewriting depth weight is calculated based on their specific similarity values, and the rewriting contribution score is obtained by weighted averaging. Structural Reorganization Contribution: Analyze the order of occurrence of segments in the user's text that are semantically related to the AI ​​content. If the order differs significantly from that in the original AI response, a structural reorganization contribution score is obtained. Finally, the three sub-scores are weighted and combined to obtain a total human semantic contribution score between 0 and 100.

[0027] S5: Integrate the independent creative confidence rating with the human semantic contribution to generate a human creative contribution index; specifically, generating the human creative contribution index includes: S51: Convert the independent creative confidence rating into a behavioral score; map the three levels of confidence rating ("high, medium, and low") to a preset base score range (e.g., high: 80-90 points, medium: 50-80 points, low: 0-50 points), and fine-tune it within this range according to the specific abnormality value calculated by S31 to obtain the final behavioral dimension score. S52: Dynamically allocate the fusion weights of the behavioral and semantic dimensions based on confidence ratings and semantic analysis confidence levels; establish a weight allocation table, with the basic principle being that the dimension with higher input data confidence levels receives a greater weight. For example, when the behavioral confidence rating is "high" and the semantic analysis is based on high-confidence AI dialogue, their weights may be close (e.g., behavioral 0.55, semantic 0.45); if the behavioral confidence rating is "low" but the semantic analysis confidence level is high, the semantic weight will increase significantly (e.g., behavioral 0.30, semantic 0.70). S53: Cross-validation correction based on the correlation features of behavior and semantics; establish a correction rule base to add or deduct points for special scenarios where behavioral evidence and semantic evidence are inconsistent. For example, when semantic evidence shows that a large amount of AI content is retained (high similarity), but behavioral evidence shows that the user has made deep modifications to this content (high post-paste modification rate), the system determines this as a valuable "critical rewriting" and adds points for positive correction. S54: A human creative contribution index ranging from 0.0 to 100.0 is obtained through weighted fusion and correction. It is calculated using the formula HAI = (behavioral weight × behavioral score + semantic weight × semantic contribution) + cross-validation correction score, with the result constrained between 0.0 and 100.0. This index intuitively reflects the degree of human dominance in this creative process. S6: The contribution statement summary, including the human contribution index, is stored on the blockchain. After the system generates the final evaluation report, the core summary information (such as HAI value, key feature hash, and timestamp) is extracted to generate a contribution statement summary. The hash value of this summary is calculated and uploaded to a designated blockchain network (such as a consortium blockchain or a public testnet). No original text or dialogue content is uploaded during this process. The report will include the transaction ID of this storage, which any third party can use to query and verify the completeness and authenticity of the report summary on the blockchain, thereby achieving credible and tamper-proof storage of the evaluation results.

[0028] Reference Figure 2 This is a block diagram of the content creation contribution evaluation system based on keyboard behavior and AI interaction proposed in this invention. The evaluation system includes: The local editing behavior collection module is configured to collect data on keyboard input rhythm, backtracking rate, and paste block distribution during the user's creation process. The AI ​​dialogue interception and caching module is configured to capture or import user interaction content with AI in real time. The behavioral fingerprint generation and independent thinking index calculation module is configured to generate an independent thinking index based on behavioral characteristics. The AI ​​text semantic comparison and contribution calculation module is configured to compare user text with AI interaction content semantically and calculate semantic contribution. The dual-engine fusion module is configured to fuse behavioral indices and semantic contributions to generate a human creative contribution index. The blockchain evidence storage module is configured to upload contribution statement summaries to the blockchain.

[0029] In one embodiment, the local editing behavior acquisition module includes: A keyboard input rhythm acquisition unit is used to statistically analyze input time intervals based on a sliding window mechanism, calculating the standard deviation and long pause ratio. Specifically, a thread-safe circular queue is used to store timestamps. The queue has a fixed length of 400 entries. Each valid keyboard input records the current timestamp and calculates the interval between the current and previous inputs, storing this interval in the queue. When the queue is full, the oldest interval is removed and the newest interval is added, implementing the sliding window mechanism. An independent background thread performs statistical analysis on the 400 intervals in the current queue every 8 seconds, outputting rhythm characteristics: the standard deviation σ reflects the degree of input rhythm fluctuation, with a preset human contribution operation σ≥500ms and simple copying σ≤100ms; the long pause ratio R_long has a preset long pause threshold of 1500ms, calculated using the formula "ΔT". n Number of times ≥1500ms / 400", with preset human cognitive operation R_long≥20% and mechanical operation R_long≤5%; rhythm fluctuation trend, calculate the slope of σ change over the last 3 statistical windows to determine rhythm stability; A rollback rate calculation unit is used to distinguish between type A and type B deletions, calculate the rollback rate, and associate it with the document structure. In specific implementation, a monotonically increasing global operation event ID is designed, and all creation operations generate standardized event entities. Through the "ContentSource" field of the event, combined with the text source corresponding to the cursor position of the deletion operation, two types of deletion behaviors are distinguished: type A deletion is the deletion of content from the Input source, which is determined as a valid rollback calculation item; type B deletion is the deletion of content from the Block source, which is not included in the rollback rate calculation and is recorded separately as a "block content clearing event". The rollback rate calculation formula is: Rollback rate (%) = (ΣDel_A / ΣAdd_Input) × 100%. The document structure elements are identified through regular expression matching, and combined with the CursorPos context of the deletion operation, structural level rollback and semantic unit rollback are distinguished. The paragraph creation cycle is divided into three stages: early stage, middle stage, and late stage. The proportion of rollback characters in each stage is counted, and three time patterns are output: concentrated rollback in the early stage, uniform and continuous rollback, and concentrated rollback in the late stage. The paste block distribution recognition unit is used to identify paste events and distinguish between AI-sourced and non-AI-sourced events. In specific implementation, when the editor gains focus, or when a paste keyboard event or right-click paste menu is detected, the system clipboard text is immediately read, a SHA256 hash and a text snapshot are generated, and stored in a temporary cache. The clipboard's dataChanged signal is monitored, and if the editor is in focus, the cached content is updated synchronously. When the number of characters increases by ≥30 or a drag-and-drop event is triggered, the newly added text is compared with the cached clipboard text: if the text overlap is ≥90%, regardless of the time interval, it is directly determined as "confirm paste" and the paste source is marked; if there is no matching content in the clipboard cache, it enters the comprehensive feature evaluation; for the collected paste operation data, the following features are calculated in real time: paste operation frequency, block size distribution, and keyboard activity after the operation.

[0030] In one embodiment, the AI ​​dialogue interception and caching module includes: A browser extension unit is used to extract dialogue content in real time from a specified AI webpage. In practice, the extension is developed to be compatible only with mainstream browsers such as Chrome, Edge, and Firefox, with permissions strictly limited to "access to a specific AI webpage domain + page DOM reading + local communication". When a user accesses the target AI webpage, the extension listens to changes in the page DOM, identifies the tag features of "user input" and "AI response", and extracts the dialogue text in the browser's memory. A WebSocket secure communication channel is used to synchronize the extracted dialogue in real time to the encrypted database of the local editor, which is structured as "timestamp + role + text + source domain + confidence level". A new "disconnection and reconnection" mechanism is added: if the communication between the browser and the local editor is interrupted, the extension temporarily stores the dialogue in the browser's local storage and automatically re-uploads it after the connection is restored. The file import unit is used to parse AI dialogue files uploaded by users. In practice, it supports the official export formats of mainstream AI platforms, such as ChatGPT's JSON export file, Claude's TXT dialogue records, and Wenxin Yiyan's HTML report. The system has a built-in dedicated parser that automatically extracts "user prompt", "AI response", "timestamp" and "model name". After parsing, a "dialogue preview confirmation page" pops up, showing the extracted dialogue list. Users can correct recognition errors and mark "medium confidence" after confirming that there are no errors, and store it in the database. A confidence leveling unit is used to assign confidence weights to AI content from different sources. In specific implementation, the confidence weights for different sources are clearly defined: 1.0 for extended records, 0.6 for official exports, and 0.3 for OCR imports. All dialogue records are uniformly stored in a "dialogue package" structure, including: dialogue ID, collection source, confidence score, role, text content, timestamp, and associated file path. AI response text is split into short sentences of 50-100 characters, and sentence vectors are generated using a sentence vector model to build an index library of "short sentence-vector-dialogue ID-confidence".

[0031] In one embodiment, the behavioral fingerprint generation and independent thinking index calculation module includes: Anomaly calculation unit; used to calculate rhythm anomaly and regression anomaly based on a preset average baseline; in specific implementation, based on multi-scenario creation sample statistics, a fixed average baseline is set: input rhythm standard deviation average baseline σ_avg=150ms, regression rate average baseline R_avg=0.15; rhythm anomaly = σ ÷ σ_avg of the current sliding window, regression anomaly = R ÷ R_avg of the current sliding window, both rounded to two decimal places; unified interpretation rules: <0.8 indicates significantly mechanical behavior, 0.8-1.2 indicates behavior close to the normal state of ordinary creators, 1.2-1.5 indicates a slight increase in thinking intensity, and >1.5 indicates a significant increase in thinking intensity; The paste pattern diagnostic unit is used to determine the paste pattern based on the AI ​​matching rate and modification rate. In practice, it outputs five structured factual data points: total number of pastes, percentage of high-confidence AI matches, average post-paste modification rate, maximum single-paste character count, and percentage of non-AI pastes. Three clear diagnostic rules are set: Pattern A is direct AI reuse: high-confidence AI matching percentage > 60% and average post-paste modification rate < 20%, diagnosed as "A large amount of unmodified high-confidence AI content was detected, with weak independent creation characteristics"; Pattern B is AI critical modification: high-confidence AI matching percentage 20%-60% and average post-paste modification rate > 50%, diagnosed as "Deep critical modification of AI content was detected, reflecting proactive thinking"; Pattern C is non-AI paste: high-confidence AI matching percentage < 20%, diagnosed as "Pasted content is mainly from non-AI channels, with low interference to independent creation evaluation." Confidence rating unit; used to output three levels of independent creative confidence rating: high, medium, and low. In specific implementation, it combines "paste mode" and "abnormality" to output three levels of confidence rating: low confidence is to meet "mode A", or rhythm abnormality < 0.8 and regression abnormality < 0.8; high confidence is to meet "mode B", and rhythm abnormality > 1.2 or regression abnormality > 1.2; medium confidence is to not meet the low / high confidence conditions.

[0032] In one embodiment, the AI ​​text semantic comparison and contribution calculation module includes: Text segmentation units; used to segment text based on semantic boundaries; in specific implementation, segmentation adopts "semantic unit priority + maximum length limit": segmentation is prioritized according to semantic boundaries, Chinese is identified by "。!?;\n", English by ".?!;\n", and code by ";{}()\n"; maximum length catch-all: if a single semantic block exceeds 1000 characters, it is truncated at the nearest comma; supports multilingual documents: automatically determines the paragraph language through character encoding and language recognition library, and matches the corresponding segmentation rules; Sentence vector matching unit; used to calculate the similarity between user text and AI content; in specific implementation, for short texts with less than 50 characters, the "context expansion" strategy is adopted, and the content of 100 characters before and after the short text is concatenated to generate sentence vectors; the similarity matching is changed from "bidirectional maximum matching" to "bidirectional Top-2 similarity weighting", that is, the average of the top two maximum similarities between the AI ​​fragment and the user text is taken, and the average of the top two maximum similarities between the user text and the AI ​​fragment is taken, and then the retention rate is calculated; A contribution-level stratified calculation unit is used to calculate original contributions, rewriting contributions, and structural reorganization contributions. In practice, based on statistics of AI-generated and human-modified text samples, two levels of similarity thresholds are set: a high similarity threshold (sim ≥ 0.90) corresponds to "AI content almost unmodified and retained"; a medium similarity threshold (0.70 ≤ sim < 0.90) corresponds to "AI content partially retained"; and a low similarity threshold (sim < 0.70) corresponds to "AI content very little retained". The formula for calculating the direct retention rate of AI text is: [Σ(char_len(A i )×W_forward(i))] / Σchar_len(A i )×100, where W_forward(i) is the AI ​​fragment A i Weights: If sim_forward(i)≥0.90, then W=1.0; if 0.70≤sim_forward(i)<0.90, then W=0.5; if sim_forward(i)<0.70, then W=0. The formula for calculating human semantic contribution is: Human semantic contribution = 30 × original contribution score + 40 × rewriting contribution score + 30 × structural reorganization contribution score. The original contribution score is calculated by multiplying the percentage of characters in document paragraphs with sim_backward(j)<0.70 by 100. The rewriting contribution score is calculated by multiplying the percentage of document paragraphs with sim_backward(j)<0.90 by 0.70. The rewriting weight of a single paragraph is calculated as "(0.90-sim_backward(j)) / 0.20", and then multiplied by 100 × 0.4 after weighted averaging. The structural reorganization contribution score is calculated as follows: if the difference between the document paragraph order and the order of the matched AI segment is ≥60%, then 30 points are awarded; if the difference is 30%-60%, then 15 points are awarded; and if the difference is <30%, then 0 points are awarded.

[0033] In one embodiment, the dual-engine fusion module includes: A dynamic weight calculation unit is used to dynamically adjust the fusion weights based on behavioral confidence ratings and semantic confidence levels. Specifically, the fusion weights are dynamically adjusted based on both behavioral confidence ratings and semantic confidence levels: when the behavioral confidence rating is high and the semantic confidence level is high, the behavioral weight is 0.55 and the semantic weight is 0.45; when the behavioral confidence rating is high and the semantic confidence level is medium / low, the behavioral weight is 0.65 and the semantic weight is 0.35; when the behavioral confidence rating is medium and the semantic confidence level is low, the fusion weight is high. When the behavioral confidence level is high, the behavioral weight is 0.45 and the semantic weight is 0.55; when the behavioral confidence level is medium and the semantic assessment confidence level is medium / low, the behavioral weight is 0.50 and the semantic weight is 0.50; when the behavioral confidence level is low and the semantic assessment confidence level is high, the behavioral weight is 0.30 and the semantic weight is 0.70; when the behavioral confidence level is low and the semantic assessment confidence level is medium / low, the behavioral weight is 0.40 and the semantic weight is 0.40. The remaining 0.2 is marked as "assessment uncertainty". Cross-validation unit; used for positive and negative correction based on behavior-semantic association tags; in specific implementation, for the "behavior-semantic reverse matching" scenario, a correction coefficient of ±5 to 20 points is added: positive correction is +15 points when the semantic-behavior association tag is "high similarity + high modification rate" and the paste mode is B; negative correction is -10 points when the semantic-behavior association tag is "low similarity + low modification rate" and the anomaly score is <0.8; no correction is applied to other scenarios; The HAI (Human Creative Contribution Index) calculation unit is used to generate a human creative contribution index ranging from 0.0 to 100.0. Specifically, the HAI calculation formula is: HAI = (W1 × HCI + W2 × S) + cross-validation coefficient, where W1 is the behavioral weight, HCI is the behavioral score, W2 is the semantic weight, and S is the human semantic contribution. Based on multi-scenario creative sample clustering analysis, the HAI interval boundaries and quantitative descriptions are determined: 90.0–100.0 indicates a high degree of consistency with independent human creation characteristics, judged by S≥85, HCI≥85, and AI high similarity ratio <5%; 75.0–89.9 indicates human-led creation with AI playing only an auxiliary role, judged by S≥70. The criteria for determining the characteristics of AI-generated features are as follows: S≥70, HCI≥50, AI high similarity percentage <20%; 50.0–74.9 indicates deep AI involvement with significant human modifications, judged by S≥50, HCI≥50, AI high similarity percentage 20%–50% and modification rate ≥40%; 25.0–49.9 indicates AI dominance with only partial or formal human modifications, judged by S<50, HCI<60, AI high similarity percentage 50%–80% and modification rate <30%; 0.0–24.9 indicates features highly consistent with AI-generated features after slight modifications or direct use, judged by S<30, HCI<40, AI high similarity percentage ≥80% and modification rate <20%.

[0034] In one embodiment, the blockchain evidence storage module uses a consortium blockchain or a public testnet, uploading only the hash value of the contribution statement summary, and supports third-party verification via transaction ID. In specific implementation, a "dual-chain backup + hierarchical evidence storage" scheme is adopted: ordinary users are connected to the Polygon Amoy testnet by default, while institutional users can switch to education / government consortium blockchains; the evidence storage content includes the HAI score, the overall SHA256 hash of the report, and the creation timestamp; the report includes an evidence storage number and a query link at the end, and third parties can verify hash consistency by entering the number.

[0035] In one embodiment, the collaborative workflow of the various modules of this system is as follows: When a user starts creating content using the local editor, the local editing behavior collection module begins to work, collecting keyboard input events, deletion events, and paste events in real time. At the same time, the AI ​​dialogue interception and caching module captures the user's interactive dialogue with AI services such as ChatGPT in real time through browser extensions, or obtains historical dialogue records through file import. After a user completes their creation and submits the text, the system initiates an analysis process: the behavior fingerprint generation and independent thinking index calculation module analyzes the collected behavior data, calculates rhythm anomaly and regression anomaly, diagnoses the pasting pattern, and outputs an independent creation confidence rating; the AI ​​text semantic comparison and contribution calculation module performs semantic comparison between the user's text and the cached AI dialogue content, and calculates the direct retention rate of AI text and the human semantic contribution. The dual-engine fusion module receives the output results from the two engines, dynamically allocates weights based on behavioral confidence rating and semantic assessment confidence, and calculates the final human creative contribution index by combining cross-validation rules. Finally, the blockchain notarization module generates a contribution statement summary from key information such as HAI value and report hash, and uploads the summary hash value to the blockchain network.

[0036] In one embodiment, the specific implementation of local editing behavior collection includes the following steps: Keyboard Input Rhythm Acquisition: Upon system startup, the user authorization process is completed, informing them that keyboard input rhythm, pasting, and other operations will be collected. Data is stored locally only. After the user agrees, the editor built on QTextEdit is launched. A dual-track acquisition system of "keyboard input stream + block operation" is adopted. Keyboard input stream acquisition is bound to the contentsChange signal. Only when a character is added and it is not undo / redo or triggered by a block operation is it considered a valid keyboard input operation. A thread-safe circular queue is used to store timestamps. The queue has a fixed length of 400 entries. Each valid keyboard input is recorded with the current timestamp and the interval between the current input and the previous input is calculated. An independent background thread performs statistics on the 400 intervals in the current queue every 8 seconds, outputting the standard deviation σ, the proportion of long pauses R_long, and the rhythm fluctuation trend. Rollback rate calculation: Based on the global operation event flow architecture, a monotonically increasing global operation event ID is designed. All creation operations generate standardized event entities. By using the "ContentSource" field of the event and combining the text source corresponding to the cursor position of the deletion operation, A-type deletion and B-type deletion are distinguished. The rollback rate calculation formula is: Rollback rate (%) = (ΣDel_A / ΣAdd_Input) × 100%. By using regular expression matching to identify document structure elements and combining the context of the deletion operation, structural level rollback and semantic unit rollback are distinguished. The paragraph creation cycle is divided into three stages: early stage, middle stage and late stage. The proportion of rollback characters in each stage is counted, and three time patterns are output. Paste block distribution recognition: When the editor gains focus or a paste event is detected, the system clipboard text is immediately read, a SHA256 hash and a text snapshot are generated and stored in a temporary cache, and the clipboard's dataChanged signal is listened for. If the editor is in focus, the cached content is updated synchronously. When the number of characters increases by ≥30 or a drag-and-drop event is triggered, the newly added text is compared with the cached clipboard text: if the text overlap is ≥90%, it is directly determined as "confirm paste"; if there is no matching content in the clipboard cache, it enters the comprehensive feature evaluation, and the pasting operation data is calculated in real time for features such as pasting operation frequency, block size distribution, and keyboard activity after the operation.

[0037] In one embodiment, the specific implementation of AI dialogue interception and caching includes the following components: Browser extension component: Develop an extension that is only compatible with mainstream browsers such as Chrome, Edge, and Firefox. The permissions are strictly limited to "access to specific AI webpage domains + page DOM reading + local communication". When a user visits the target AI webpage, the extension listens to changes in the page DOM, identifies the tag features of the dialogue bubble, extracts the dialogue text, and uses a WebSocket secure communication channel to synchronize the structured dialogue data to the encrypted database of the local editor in real time. A new "disconnection and reconnection" mechanism has been added to ensure data integrity. File import component: Supports parsing official export formats of mainstream AI platforms. Built-in dedicated parser automatically extracts dialogue content. After parsing, a "dialogue preview confirmation page" pops up, allowing users to correct recognition errors. After confirmation, confidence is marked and stored in the database. For screenshot files, the local Tesseract OCR engine is called to extract text, and then the text structure is intelligently analyzed through a pre-built local lightweight model to distinguish between user input and AI response. Confidence Management System: All dialogue records are uniformly stored in a "dialogue package" structure, which includes fields such as dialogue ID, collection source, and confidence score. AI response text is split into short sentences, sentence vectors are generated, an index library is built, and the confidence weight of different sources is clearly defined: extended records are 1.0, official exports are 0.6, and OCR imports are 0.3. When subsequent modules call dialogue data, they need to calculate the comprehensive confidence score according to the weight.

[0038] In one embodiment, the specific implementation of dual-engine fusion includes the following steps: Input data processing: Receives behavioral confidence rating, rhythm anomaly, regression anomaly, and paste pattern label from the behavior engine; receives human semantic contribution, semantic evaluation confidence, and semantic-behavior association label from the semantic engine; Behavioral score conversion: The behavioral confidence rating is mapped to a base score range: high rating corresponds to 80.0-90.0, medium rating corresponds to 50.0-79.9, and low rating corresponds to 0.0-49.9. Fine-tuning is then applied to the base score based on the degree of anomalousness: if the rating is high and any anomalousness is >1.5, +5 points are awarded; if the rating is medium and any anomalousness is >1.5, +2 points are awarded; if the rating is low and the paste mode is B, +10 points are awarded. Dynamic weight calculation: The fusion weight is determined by looking up a table based on the behavioral confidence rating and semantic assessment confidence. For example, when the behavioral confidence rating is high and the semantic assessment confidence is high, the behavioral weight is 0.55 and the semantic weight is 0.45; when the behavioral confidence rating is low and the semantic assessment confidence is high, the behavioral weight is 0.30 and the semantic weight is 0.70. Cross-validation correction: Corrections are made for specific scenarios; when the semantic-behavioral association label is "high similarity + high modification rate" and the paste mode is B, a positive correction of +15 points is made; when the semantic-behavioral association label is "low similarity + low modification rate" and the anomaly score is <0.8, a negative correction of -10 points is made. HAI Calculation and Output: The final score is calculated according to the formula HAI=(W1×HCI+W2×S)+cross-validation coefficient. Based on the preset interval in which the HAI value falls, the corresponding contribution feature conclusion is output, and at the same time, a visual chart is generated: HAI sub-item radar chart shows HCI, S and sub-item scores; weight percentage pie chart shows the weight distribution of this fusion; similarity distribution heat map annotates the similarity between each paragraph of the document and the AI ​​segment.

[0039] In one embodiment, the blockchain evidence storage module adopts a tiered evidence storage scheme: ordinary users are connected to the PolygonAmoy testnet by default, while institutional users can switch to education / government consortium blockchains. The stored content only includes the HAI score, the overall SHA256 hash of the report, and a summary of the creation timestamp. After the system generates the evidence storage transaction, it attaches an evidence storage number and a query link at the end of the report. Third-party verifiers can verify hash consistency by entering the evidence storage number, confirming that the report has not been tampered with since its generation; all raw data is processed locally, without uploading any creation content or conversation records, thus protecting user privacy while achieving reliable verification of the evaluation results.

[0040] In one embodiment, the various modules of the system work collaboratively through preset linkage rules, as shown in the following example: Data transfer between modules: The local editing behavior acquisition module outputs a structured behavior data package, including keyboard input rhythm features, block operation feature data, and their mapping relationship with text paragraphs. The AI ​​dialogue interception and caching module outputs a weighted dialogue evidence package, including AI response phrases, sentence vectors, dialogue IDs, acquisition sources, and confidence scores.

[0041] Behavioral-semantic cross-validation rules: Rule 1: If a paragraph has "AI source pasting ratio ≥ 60%", "post-paste modification rate ≥ 50%", and "semantic analysis shows originality ≥ 70%", it is judged as a "critical innovation contribution" and its weight in the contribution evaluation is increased by 20%.

[0042] Rule 2: If "unverified modification of external content accounts for ≥30%" and "AI source clarity is ≥80%", it is judged as "potential AI dependence", and the contribution assessment tends to be conservative.

[0043] Rule 3: A combination of high-frequency fluctuations (σ≥500ms) and high retracement rate (15%-30%) is considered a high-intensity critical modification signal.

[0044] Rule 4: Low-rhythm fluctuations (σ≤100ms) + low backoff rate (≤5%) + block operation frequency ≥3 times / minute linkage are judged as direct adoption of AI content signals.

[0045] Confidence linkage rule: When performing semantic comparison, sentence vectors of high-confidence AI dialogue segments are used first, and low-confidence segments are only used as auxiliary references. When module 1.3 matches pasted text, only when the confidence of the matched AI segment is ≥60% will it be included in the calculation of "AI source pasting ratio".

[0046] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0047] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0048] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for evaluating the creative contribution of content based on keyboard behavior and AI interaction, characterized in that, Includes the following steps: Collect data on keyboard input rhythm, backtracking behavior, and paste operation during the user's creation process; Obtain the content of the interaction between the user and the AI ​​model; An independent creative confidence rating is generated based on the analysis of collected behavioral data. The semantics of user-generated text are compared with the acquired AI interactive content to calculate the human semantic contribution. By integrating the independent creative confidence rating with the human semantic contribution, a human creative contribution index is generated; The summary of contribution statements, which includes an index of human creative contribution, will be stored on the blockchain.

2. The method for evaluating the contribution of content creation based on keyboard behavior and AI interaction as described in claim 1, characterized in that, The collected keyboard input rhythm includes: A sliding window mechanism is used to record the time interval of continuous keyboard input; The standard deviation of the input rhythm and the proportion of long pauses are calculated based on time intervals. Based on a preset rhythmic characteristic baseline, determine whether the input rhythm is a human-contributed operation or a mechanical operation.

3. The method for evaluating the contribution of content creation based on keyboard behavior and AI interaction as described in claim 1, characterized in that, The data rollback behavior includes: Distinguish between deleting keyboard input and deleting pasted content; The effective rollback rate is calculated based on the deletion of keyboard input. Analyze the relationship between fallback behavior and document structure to identify structural hierarchy fallback and semantic unit fallback.

4. The method for evaluating the creative contribution of content based on keyboard behavior and AI interaction as described in claim 1, characterized in that, The data collection and pasting operation includes: The paste operation is determined by pre-caching the clipboard content and comparing it with the text. Distinguish between AI-generated and non-AI-generated paste; Calculate the post-paste modification rate, which is the proportion of pasted content that is subsequently deleted.

5. The method for evaluating the creative contribution of content based on keyboard behavior and AI interaction as described in claim 1, characterized in that, The acquisition of the interactive dialogue content between the user and the AI ​​model includes: Extract conversation content in real time from a designated AI service webpage using a browser extension; Supports importing official exported dialogue files from AI platforms; OCR recognition is used to extract the dialogue text from screenshots of conversations. Assign credibility weights to dialogue content from different sources.

6. The method for evaluating the creative contribution of content based on keyboard behavior and AI interaction as described in claim 1, characterized in that, The generated independent creative confidence rating includes: The rhythm anomaly and regression anomaly are calculated based on the preset average behavior baseline; The pasting mode is determined based on the AI ​​content matching rate and the post-paste modification rate, including AI direct reuse mode, AI critical modification mode and non-AI pasting mode. Based on the anomaly calculation results and the paste mode judgment results, a three-level independent creation confidence rating of high, medium and low is output.

7. The method for evaluating the contribution of content creation based on keyboard behavior and AI interaction as described in claim 1, characterized in that, The calculation of human semantic contribution includes: Intelligent text segmentation based on semantic boundaries; Convert text fragments into vector representations and calculate similarity; Set a similarity threshold to distinguish between content with high similarity, medium similarity, and low similarity; Based on the similarity calculation results, the human semantic contribution is calculated from three dimensions: original contribution, rewriting contribution, and structural reorganization contribution.

8. The method for evaluating the creative contribution of content based on keyboard behavior and AI interaction as described in claim 1, characterized in that, The generated human creative contribution index includes: Transform independent creative confidence ratings into behavioral scores; The fusion weights of the behavioral and semantic dimensions are dynamically assigned based on confidence ratings and semantic analysis confidence levels. Cross-validation correction is performed based on the correlation features between behavior and semantics; The human creative contribution index, ranging from 0.0 to 100.0, is obtained through weighted fusion and correction.

9. A system for evaluating the creative contribution of content based on keyboard behavior and AI interaction, characterized in that, include: The local editing behavior collection module is configured to collect data on keyboard input rhythm, backtracking rate, and paste block distribution during the user's creation process. The AI ​​dialogue interception and caching module is configured to capture or import user interaction content with AI in real time. The behavioral fingerprint generation and independent thinking index calculation module is configured to generate an independent thinking index based on behavioral characteristics. The AI ​​text semantic comparison and contribution calculation module is configured to compare user text with AI interaction content semantically and calculate semantic contribution. The dual-engine fusion module is configured to fuse behavioral indices and semantic contributions to generate a human creative contribution index. The blockchain evidence storage module is configured to upload contribution statement summaries to the blockchain.

10. The content creation contribution evaluation system based on keyboard behavior and AI interaction as described in claim 9, characterized in that: The local editing behavior acquisition module includes a keyboard input rhythm acquisition unit, a rollback rate calculation unit, and a paste block distribution recognition unit; The AI ​​dialogue interception and caching module includes a browser extension unit, a file import unit, and a confidence level unit; The behavioral fingerprint generation and independent thinking index calculation module includes an anomaly calculation unit, a pasting pattern diagnosis unit, and a confidence rating unit. The AI ​​text semantic comparison and contribution calculation module includes a text segmentation unit, a sentence vector matching unit, and a contribution hierarchical calculation unit. The dual-engine fusion module includes a dynamic weight calculation unit, a cross-validation unit, and a human creative contribution index calculation unit. The blockchain evidence storage module supports uploading the hash value of the contribution statement summary to the consortium blockchain or public test chain.

Images