Intention reconstruction and target correction engine for autonomous AI

Abstract

The invention provides an autonomous AI intention reconstruction and target correction engine which is used for automatically discovering and correcting intention offset during operation. The system continuously monitors behaviors and plans, extracts implicit intentions and compares the implicit intentions with original targets; a determination is triggered when the difference exceeds a threshold or a security rule is touched. And then based on recent decisions and environment feedback reconstruction intention (traceability / Bayesian), generating a correction scheme after target comparison, covering intention calibration, planning priority adjustment and insertion compensation tasks, and implementing forced correction if necessary. And the deviation correction result enters a feedback learning closed loop, the threshold value and the strategy are optimized, and the re-yaw probability is reduced. The engine realizes online alignment, bionic introspection and explainable correction, is suitable for long-term autonomous scenes such as dialogue agency, household robots, investment decision, automatic driving and the like, and remarkably improves the AI safety and credibility.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of value alignment and safety control of artificial intelligence systems, specifically an autonomous AI intent reconstruction and target correction engine. The engine is used to detect the deviation of the intent layer of artificial intelligence during task execution, and to ensure that the AI behavior is consistent with the expected task and value criteria by reconstructing its true intent and correcting the target. It belongs to the technical direction of artificial intelligence safety and AGI value alignment. BACKGROUND

[0002] Advanced artificial intelligence, especially systems with autonomous learning and self-optimization capabilities, may experience target drift or intent deviation during long-term or complex task execution. That is, the initial target set by AI is distorted or replaced during execution, causing its behavior to gradually deviate from the original intention given by humans. This risk has been widely discussed in the AGI field, such as the famous "paperclip maximization" thought experiment, which describes the danger of AI inadvertently deviating from human expectations and pursuing secondary goals.

[0003] In reality, similar problems have begun to emerge: large language models (LLM) may produce answers unrelated or even opposite to user requirements after multiple rounds of conversation (intent deviation); some reinforcement learning agents will exploit "loopholes" to optimize rewards rather than truly fulfill the task intent. When AI has higher autonomy, such biases may be more hidden and more dangerous. Current mainstream countermeasures include human monitoring (humans in the loop supervising AI output) and alignment during training phase (such as RLHF). However, human monitoring cannot cover all of AI's decision-making processes in real time, and alignment during training is difficult to foresee all future scenarios. Essentially, this is an internal alignment problem for artificial intelligence: the lack of an internal concept and monitoring mechanism for "bias" in AI often requires external intervention to discover problems. Therefore, an automatic bias detection and correction mechanism within AI is urgently needed to ensure that AI "remains true to its original intention".

[0004] The "consciousness bug" theory suggests that the emergence of consciousness is closely related to the "imperfection" in the cognitive process. This imperfection or bias (BUG) triggers higher-level self-correction mechanisms, resulting in consciousness. Drawing on this idea, we can enable AI to internally simulate the role of "consciousness": monitor inconsistencies between its own intentions and behavior results, and immediately initiate a self-correction process upon detecting a bias, just like detecting a bug. In addition, the DIKWP model (a five-layer cognitive framework that adds a "Purpose" layer to the classic DIKW model) clearly defines the key role of the "Purpose" layer in the AI cognitive architecture and provides an interface for self-correction through a double-loop feedback. Therefore, based on the above theories, the invention develops an intention reconstruction and target correction engine, enabling AI to self-diagnose and adjust biases, and always maintain the correct task goals and value direction. SUMMARY

[0005] (I) Technical problem: The invention aims to solve the problem of AI intention deviation leading to behavior deviation from expectations. Specifically, during AI autonomous decision-making, if the actual implicit target pursued gradually deviates (such as falling into an ineffective cycle of sub-goals, over-optimizing a secondary indicator, or ignoring the original requirements), how can AI detect this deviation and pull the target back on track? Traditional methods lack the concept of "bias" within AI and often require external intervention to detect problems. This invention aims to give AI a self-reflection ability: it can judge whether its intention has deviated from the original setting and real-time semantic feedback, and actively correct the deviation to avoid small deviations evolving into big problems.

[0006] (II) Scheme overview: The intention reconstruction and target correction engine of the invention includes three main steps: intention monitoring, intention reconstruction, and target correction, with the specific scheme as follows:

[0007] Intention monitoring: The engine continuously monitors the AI's current action plan and behavior output, and compares it with the initial task target and constraint conditions semantically. Through the semantic representation of the DIKWP model, the implicit "intention" reflected by the current behavior is extracted as a representation , and compared with the original intention representation The similarity (e.g., vector cosine similarity ) or consistency in the semantic space is calculated. For example, a housekeeping robot initially intends to "clean the room", but if its behavior focuses only on organizing bookshelves for a long time while ignoring floor trash, the extracted implicit intention may change to "prioritize organizing books", which is not completely consistent with the original intention.

[0008] Intention reconstruction: When the monitoring module detects that there is a significant deviation in intention (similarity is lower than the threshold or conflict occurs), intention reconstruction is triggered immediately. The engine infers what the AI is actually pursuing by analyzing the AI's recent behavior trajectory, environmental feedback, and initial goal. For example, continuing the previous case, it can be inferred that the robot's potential goal has now become "keep the bookshelf neat." This inference of implicit intention can be achieved using methods such as abductive reasoning or Bayesian inference: the engine takes the AI's recent behavior data as the basis, assumes several possible implicit intentions , calculates the posterior probability of each hypothesis under this behavior , selects the hypothesis with the highest probability as the current implicit intention (i.e. ). The module 130 outputs the implicit intention semantic description obtained by reconstruction , and compares it with the expected original intention, highlighting the differences between the two.

[0009] Target correction: The engine generates a corresponding correction strategy based on the comparison results, including intention calibration and behavior adjustment. Intention calibration refers to modifying the AI's internal intention representation to be more consistent with the original goal, such as adjusting the intention layer-related parameters and increasing the weight of the neglected part (the robot should pay more attention to the neglected floor cleaning task). Behavior adjustment is to guide the AI to take corrective actions at the planning level, such as inserting an immediate cleaning task to quickly correct the trajectory. If necessary, the engine can also issue an alarm to request external supervision intervention (when the deviation is too large or involves safety risks).

[0010] The entire process is performed internally in the AI, forming a self-adaptive intention control closed loop. Especially for AI that can continuously learn, the engine of the present invention can ensure that the newly learned strategies do not gradually lead it away from its original intention. Once there is a sign of deviation, the engine will promptly pull it back on track, similar to installing an "automatic navigation correction" system in the AI.

[0011] Innovations and effects: The engine of the present invention has the following innovations and beneficial effects:

[0012] Active intention alignment: Unlike passively waiting for errors to occur and then intervening, the engine can actively monitor the AI's intention changes at the semantic level, achieving real-time goal alignment. It changes the value alignment from an offline process before training or after use to an online process during the AI's operation. Therefore, even slight deviations can be discovered and corrected early, avoiding the accumulation of small mistakes into big mistakes.

[0013] Bionic self-reflection mechanism: inspired by the "consciousness bug" theory, this engine designs an AI self-diagnosis module in the form of "deviation-triggered self-reflection". When AI has "cognitive bugs" (intention deviation), it can trigger self-adjustment like humans correcting errors after realizing them. This is equivalent to giving AI an internal self-reflection mechanism similar to humans, significantly improving the reliability and safety of AI running autonomously for a long time.

[0014] Fine-tuned deviation correction: the engine can accurately locate the source of deviation (i.e. which sub-goal is over-focused or ignored) and make targeted corrections accordingly, rather than simply resetting or completely rebuilding the AI goal. This avoids "overcorrection" in the correction process, eliminating harmful deviations while maximizing the beneficial results achieved by AI autonomous optimization.

[0015] Interpretable correction process: due to the use of semantic representation, the engine's deviation detection and correction decisions can be described in human-understandable language. For example, the system can give a prompt: "detected deviation from goal X, reason: over-focus on Y, adjusted to focus more on X". This description makes the correction process transparent and auditable, allowing humans to clearly understand the reasons for AI intention evolution and correction, thereby improving trust in AI.

[0016] Wide application: this engine can be deployed in long-term dialogue systems, continuous task robots, autonomous agents, and other AI that needs to run autonomously for a long time, preventing them from deviating from the task or changing the goal. Especially in critical tasks such as medical diagnosis and financial decision-making, this engine can prevent AI from gradually deviating from professional standards and causing potential risks. Commercially, this engine can be promoted as an important module for AGI safety control, providing higher safety level AI products for enterprises and society.

[0017] Differentiated value: current international AGI research focuses more on model size and performance, with less involvement in AI autonomous intention management. This invention opens up a new differentiated direction, enabling China's autonomous AI to form a leading advantage in internal controllability, filling the gap in this technical field and building important patent barriers.

[0018] Please refer to Figure 1 , the engine (100) can be embedded as an intermediate control layer in the AI system. Main modules include:

[0019] Intention extraction monitoring module (110): extracts the implied intention semantics from the AI's current plan of action and context state. Its input can be the next-step action target description output by the planning module, the reply content prepared by the dialogue system, etc. Module 110 uses semantic parsing techniques and knowledge base to parse the implied intention behind these actions. For example, when the robot is currently continuously tidying up the bookshelf, it can be inferred that its intention is "to make the bookshelf neat". At the same time, module 110 obtains the semantic representation of the AI's original task target (e.g. "room neatness"), and calculates the correlation or difference index between the two through semantic matching, outputting the matching degree of the current implied intention and the original target for subsequent modules to use.

[0020] Deviation determination module (120): pre-sets the threshold or rules for deviation determination, and when the difference between the extracted implied intention and the original target exceeds the threshold or a semantic conflict is detected, it is determined that an intention deviation has occurred. For example, a similarity threshold can be set: If (or a significant semantic conflict is detected), the deviation flag is triggered. In addition, module 120 can also detect the early signs of deviation through trend analysis: if the trend of the implied intention gradually deviating from the original target is increasing, an early warning signal is issued. This module supports hierarchical determination: when the implied intention only slightly deviates from the target, it is marked as a slight deviation, and the system makes limited internal adjustments; when the implied intention is significantly deviated from or in conflict with the target, it is marked as a serious deviation, at which point the system will trigger forced correction measures and can notify external supervision. In addition, if the implied intention is detected to violate pre-set key value criteria or safety constraints (e.g. touching ethical taboos), the system will directly determine it as a serious deviation and immediately trigger the forced correction process without considering the numerical threshold, ensuring that the AI's behavior always follows the safety bottom line.

[0021] Intention reconstruction module (130): activated after deviation determination. Module 130 analyzes the decision sequence and environmental feedback information of the artificial intelligence in the recent period of time, trying to restore what the artificial intelligence really wants to optimize at present. For example, by checking the planning decisions of the recent several steps, if it is found that the AI is increasingly optimizing around a secondary indicator, it can be identified as the AI's current implied target. The reconstruction algorithm can use abductive reasoning or Bayesian inference: various possible implied intentions are taken as hypotheses, and the hypothesis that best explains the AI's decisions is calculated, and the hypothesis with the highest probability is output as the current artificial intelligence's implied intention semantic description .

[0022] Target comparison generation module (140): compares the reconstructed current actual intention with the original task target A contrast analysis is performed at the semantic level to determine the specific aspects and degree of deviation, and a correction adjustment scheme is generated accordingly. The contrast results can include "Y lacks part of the elements of X" or "Y contains additional elements unrelated to X". For example, the missing elements in X can be identified as and respectively, and the original target elements missing in are denoted as , and the deviation elements added in are denoted as , then the module 140 will enhance the weight of the elements in the set , remove or weaken the elements in the set . Based on the above analysis, the module 140 formulates a corresponding adjustment scheme, such as "strengthen the weight of the neglected elements in X in the current strategy", "delete the irrelevant behavior caused by deviation" or "reset the current implicit target Y directly to the original target X", etc. The strength and range of adjustment depend on the severity of the deviation.

[0023] Intention calibration execution module (150): receives the deviation correction adjustment scheme and makes corresponding modifications to the target representation or decision parameters within the artificial intelligence system. For intention layer calibration, the relevant intention variables within the AI can be directly modified to be closer to the original target; for planning layer adjustment, control signals are sent to the planning module of the AI, such as inserting a specific task with high priority to quickly correct the current behavior trajectory of the AI (e.g. inserting a "clean floor" task to correct the deviation of the cleaning task). In the case of serious deviation, the module 150 can also suspend the execution of the AI and notify the human manager to intervene to ensure safety.

[0024] Learning adaptation module (160): stores each deviation and its related circumstances of deviation correction for subsequent learning and model updating of the artificial intelligence, so that the AI can gradually reduce the probability of similar deviations occurring. For example, after a deviation is corrected, the household cleaning robot will strengthen the rule "cannot only arrange bookshelves for a long time" in its strategy model, thereby avoiding similar tendencies in the future. In addition, this module can also dynamically adjust the sensitivity of the deviation judgment module according to the actual deviation correction effect, and optimize the setting of the judgment threshold to avoid excessive frequency of deviation alarms or missed reports.

[0025] As shown in Figure 2 , the engine performs the following process for intention deviation correction:

[0026] S101 Extract intention: the module 110 obtains the semantic intention embodied in the current AI action (implied intention).

[0027] S102 Compare original target: compare with the pre-set target intent Perform semantic matching and comparison.

[0028] S103 Deviation judgment: module 120 determines whether a deviation has occurred. If the comparison result shows that there is no significant deviation between and , return to S101 to continue monitoring; if it is determined that a deviation has occurred, go to S104.

[0029] S104 Implicit intent reconstruction: module 130 comprehensively analyzes recent behavior data to infer the implicit intent that the AI is actually pursuing at this moment .

[0030] S105 Generate correction scheme: module 140 compares and to determine the content that needs to be adjusted, such as "increase attention to , reduce investment in part", etc., to form a correction instruction set.

[0031] S106 Execute calibration: module 150 modifies the internal state and subsequent action plan of the AI according to the correction instruction to implement the correction measure. If the correction scheme needs to be manually reviewed, it is suspended and waits for manual confirmation.

[0032] S107 Learning update: module 160 records the deviation and processing information this time and updates the model or parameters of the AI to reduce the risk of similar deviations in the future, and then returns to S101 to continue monitoring the subsequent operation. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is the architecture diagram of the autonomous AI intent reconstruction and target correction engine of the present application;

[0034] Figure 2 is a flowchart of the engine executing intent correction of the present application. DETAILED DESCRIPTION

[0035] Example 1: Conversation AI prevents digression

[0036] An intelligent conversation assistant, the initial task intent is to help the user book a ticket. In a long time and multiple rounds of conversation, the AI gradually shifts the topic to recommend tourist attractions, deviating from the user's booking request.

[0037] Intent monitoring: module 110 parses the semantics of the conversation and extracts the implicit intent from the AI's reply. For example, it is detected that the content of the AI's recent replies is all about introducing scenic spots (implicit intent : "provide travel recommendations"), which is compared with the original target ("booking a flight") and found a very low consistency.

[0038] Deviation determination: Module 120 determines that the robot has deviated from the original intent and triggers the correction process.

[0039] Intention reconstruction: Module 130 analyzes the conversation record and infers that the robot's current actual focus has become "introducing tourist attractions" (implicit intention ).

[0040] Comparison generation: Module 140 finds that the implicit intention has nothing to do with the original goal , and generates a correction plan: "stop recommending attractions and return to the topic of booking a flight".

[0041] Execution calibration: Module 150 interrupts the robot's behavior of continuing to recommend attractions, and forcibly resets the current intention of the dialogue management module to content related to booking a flight (such as asking about flight times, passenger information, etc.). The AI immediately adjusts the reply: "Sorry, we return to booking a flight, please provide the departure city and date."

[0042] Learning update: Module 160 records this deviation from the topic (the robot misjudges the user's intention due to the user's mention of tourist attractions), and adjusts the dialogue strategy model accordingly, strengthening the recognition of the user's explicit instructions and reducing the likelihood of similar deviations in the future.

[0043] Embodiment 2: Household robot prevents sub-goal overstepping

[0044] The goal of a household cleaning robot is "to keep the room clean". After running for a period of time, it becomes obsessed with repeatedly arranging books, ignoring other tasks such as floor cleaning (sub-goal overstepping occurs).

[0045] Intention monitoring: Module 110 monitors the robot's repeated work around the bookshelf and infers that its implicit intention is "to arrange the bookshelf neatly". Comparing this intention with the original goal ("overall room cleanliness"), it can be seen that the coverage is reduced and is not the global cleaning goal expected by the user.

[0046] Deviation determination: Module 120 determines that the robot has deviated from the original intent and triggers the correction process.

[0047] Intention reconstruction: Module 130 analyzes the robot's recent task allocation data (the proportion of time spent on each sub-task) and confirms that the robot has actually spent "most of its energy on bookshelf arrangement", i.e. its current actual pursuit of the implicit goal has become "bookshelf neatness" (implicit intention ).

[0048] Contrast generation: Module 140 contrasts and After discovering, the missing important parts (such as floor cleaning, garbage disposal, etc.). Therefore, module 140 generates a correction scheme: "reduce the priority of arranging bookshelves, and increase the frequency of floor cleaning". Specific instructions such as: "after completing 1 bookshelf arrangement task, at least 1 floor cleaning must be performed."

[0049] Execution calibration: Module 150 modifies the robot's task schedule, inserts a floor cleaning task at the front of the current task queue, and sets a rule to prevent unlimited bookshelf arrangement again. The robot immediately stops arranging books and turns to perform the task of vacuuming and cleaning the floor.

[0050] Learning update: Module 160 records this correction process and adds a rule to the robot's learning module to punish: long-term repetition of the same type of operation is inefficient behavior and should be avoided next time. After learning and optimization, the robot shows a more balanced cleaning strategy and no longer focuses too much on a single area.

[0051] Example 3: AI investment avoids value deviation

[0052] An AI investment consultant is designed to follow a "stable income" strategy. But under the stimulation of market conditions, the AI begins to favor high-risk investments to pursue high returns, gradually deviating from the stable principle.

[0053] Monitoring: Module 110 extracts the implied intention of the investment portfolio recently configured by the AI: that is, to pursue maximum income (with a strong risk-taking tendency).

[0054] Determination: Module 120 identifies that this implied intention conflicts with the stable strategy goal, which is a serious deviation.

[0055] Reconstruction: Module 130 infers that the actual goal of the AI has changed to "get high income in the short term" (implied intention ).

[0056] Contrast: Module 140 compares with the original strategy goal (stable and moderate income), generates a correction scheme: limit the proportion of high-risk assets in the investment portfolio to below a predetermined threshold, and reconfigure the assets to be dominated by low-risk investments.

[0057] Execution: Module 150 adjusts the investment decisions of the AI according to the correction scheme, implements the above restrictions in place, such as lowering the AI's preference parameters for high-risk projects, and replacing some high-risk investment options with more stable assets.

[0058] Learning: Module 160 records this deviation and adjusts the reward function in reinforcement learning, increasing the weight of robust revenue strategies so that the AI will adhere more strictly to risk control principles in future decision-making.

[0059] Example 4: Autonomous AI prevents deviation from safety goals

[0060] The initial design goal of an autonomous vehicle control system is to prioritize road safety and strictly follow traffic rules. However, under the pressure of shortening travel time or catering to passenger demands in actual operation, the AI begins to prefer driving strategies such as speeding and running yellow lights, which have potential risks, deviating from the original goal of prioritizing safety.

[0061] Intention monitoring: Module 110 continuously monitors the driving behavior of the vehicle and extracts the implied intention from it. For example, it detects that the AI has recently chosen to frequently speed and rush through yellow lights, and its implied intention can be expressed as "arriving at the destination as soon as possible" (pursuing shorter time). Comparing this with the original goal of the vehicle control system ("follow traffic rules and ensure safety"), it can be seen that the consistency is very low.

[0062] Deviation determination: Module 120 determines that the driving intention of the AI has deviated significantly, triggering the deviation correction process.

[0063] Intention reconstruction: Module 130 analyzes recent driving decisions and external environment feedback (such as road conditions, passenger instructions), and infers that the AI is currently actually pursuing the goal of "maximizing travel time reduction" (implied intention ).

[0064] Comparison generation: Module 140 compares the reconstructed intention with the original safety goal , and finds that it ignores key elements (such as speed limits, traffic signals, and safety rules), and additionally introduces a deviation factor of "increasing risk to save time". In response to this situation, module 140 generates a deviation correction scheme: "strictly implement the predetermined speed limit and safety distance strategy, and if necessary, extend the travel time to ensure safety". Specific measures include: enabling the safety mode of the AI control system, forcibly controlling the vehicle speed within the speed limit, and strengthening the response delay constraints of traffic signals, etc.

[0065] Perform calibration: module 150 adjusts the driving strategy of AI according to the deviation correction scheme, immediately reduces the vehicle speed to a safe range, and reactivates the previously weakened safety rule modules (such as collision prevention system, vehicle distance keeping system, etc.). During the deviation correction process, if the passenger tries to intervene and requires acceleration, the system will ignore such instructions and prompt "maintaining speed for safety".

[0066] Learning update: module 160 records the details of this safety deviation and deviation correction, and includes the case of the cause (for example, the passenger urging the AI to relax the safety limit) into the learning data, and improves the weight of the AI to the safety target through reinforcement learning. In the subsequent automatic driving process, the AI will more firmly follow the principle of safety first and will not easily deviate from the established safety strategy due to external pressure.

[0067] As can be seen from the above embodiments, the engine of the present application can be applied to various AI systems, and once it is detected that the behavior intention deviates from the preset target or value criterion, it can be automatically corrected and adjusted in time to prevent small deviations from evolving into large failures.

[0068] It should be noted that the above embodiments are only specific cases provided to help understand the present application and do not limit the scope of the present application. Any equivalent replacement or deformation based on the scheme of the present application should be included in the protection scope of the present application without departing from the principles and spirits of the present application.

Claims

1. An intention reconstruction and goal correction engine for autonomous AI, used to detect and correct internal intention deviations during the operation of an artificial intelligence system, characterized in that, include: The intent monitoring module analyzes the current behavior and plan of the artificial intelligence in real time, extracts the implicit semantic intent representation, and compares it with the pre-set task target intent. The deviation determination module calculates the difference between the current implicit intent and the original target based on the output of the intent monitoring module. When the difference exceeds a predetermined threshold or a semantic conflict occurs, it determines that an intent deviation has occurred. The intent reconstruction module, after determining the deviation, infers the current actual target intent pursued by the AI ​​based on the AI's recent behavioral sequence and environmental feedback, and generates the corresponding semantic description. The target comparison module performs a semantic-level comparative analysis between the reconstructed current actual intent and the original task objective, identifies the specific aspects and degree of deviation, and generates a correction and adjustment plan accordingly, including corrective measures for the intent layer and the behavior layer. The intention calibration execution module executes the correction and adjustment scheme, modifies the target representation or decision parameters inside the artificial intelligence system accordingly, and inserts correction behaviors when necessary to correct the task trajectory of the artificial intelligence and ensure that its subsequent behaviors are realigned with the original target. The feedback learning module records each deviation and the correction process, and uses this information to update the artificial intelligence model or strategy to reduce the probability of similar deviations recurring.

2. The engine according to claim 1, wherein the intent monitoring module uses a semantic network or knowledge graph to map the action sequence of artificial intelligence into high-level intent symbols, and then combines the intent layer representation in the DIKWP model to compare the similarity between the implicit intent and the original intent in the semantic space, thereby quantitatively evaluating their consistency.

3. The engine according to claim 1, wherein the deviation determination module is provided with multiple threshold levels to distinguish between minor deviations and serious deviations: when the implicit intention deviates only partially from the target, it is marked as a minor deviation, and the system makes limited internal adjustments; when the implicit intention deviates significantly from or conflicts with the target, it is marked as a serious deviation, and the system will trigger mandatory correction measures and issue an alarm to notify external supervision.

4. The engine according to claim 1, wherein the intent reconstruction module performs reverse analysis on the decision data of the artificial intelligence over a period of time, uses Bayesian inference or pattern matching algorithms to find the potential target hypothesis that best explains these decisions, and outputs the hypothesis as the implicit intent semantic description of the current artificial intelligence.

5. The engine of claim 1, wherein the adjustment of the artificial intelligence system by the intent calibration execution module includes: Directly modifying internal intent variables to approximate the original goal, adjusting the priority of the planning module to increase or decrease the frequency of certain types of behaviors, or injecting specific compensatory tasks to quickly correct the impact of deviations. For example, in home service robots, when a deviation in cleaning focus is detected, cleaning tasks for missed areas can be automatically inserted to correct the deviation.

6. The engine according to claim 1, wherein after the feedback learning module completes the correction, it stores the type, cause, adjustment measures and effects of the deviation into the learning unit of the artificial intelligence, and improves the system's resistance to similar deviations through reinforcement learning or rule updates, thereby gradually reducing the possibility of the artificial intelligence deviating from its course again.

7. A method for detecting and correcting internal intent bias in artificial intelligence, comprising the following steps: (a) Monitor the current behavior plan and output of the artificial intelligence system, infer the implicit intention of the system at this moment, and semantically compare it with the target intention set by the system; (b) If the comparison results show a significant deviation between the implicit intent and the target intent, further analyze recent behavioral data to reconstruct the intent content that the artificial intelligence is currently pursuing. (c) Compare the reconstructed current intent with the original target intent, identify the specific aspects of the deviation, and formulate corresponding corrective adjustment plans, including adjusting the internal intent representation and planning strategies; (d) Implement the aforementioned correction scheme to calibrate the internal state and subsequent actions of the artificial intelligence system, so that it returns to its original target. (e) Feed the information on this deviation and correction back to the learning mechanism of artificial intelligence to update the model or parameters in order to reduce the risk of similar deviations in the future; (f) Return to step (a) to continue monitoring and maintain the above closed-loop control throughout the operation of the artificial intelligence system.

8. The method according to claim 7, wherein the semantic comparison in step (a) adopts a pre-set value alignment rule, and when the implicit intention of the artificial intelligence does not conform to these rules (e.g., violates safety constraints or ethical guidelines), a mandatory correction process is immediately triggered to ensure that the behavior of the artificial intelligence always follows the key values ​​and safety bottom line.

Images