Detection of AI-generated text
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- NEC LABORATORIES AMERICA INC
- Filing Date
- 2024-05-06
- Publication Date
- 2026-06-16
Smart Images
Figure 2026519433000001_ABST
Abstract
Claims
1. A computer implementation method for detecting artificial intelligence (AI) generated text using a processor device, The candidate text is truncated (110), and the prefix text and residual text are obtained by using the truncation module. Using the prefix text with an AI text generation model, a regenerated model text is obtained (120), The detection result is predicted by comparing the n-gram similarity between the regenerated model text and the residual text (130), A computer implementation method comprising (140) identifying whether the candidate text was AI-generated by providing explanatory text based on the detection result.
2. The computer implementation method according to claim 1, further comprising labeling candidate text in a medical record so that it can be identified as AI-generated for presentation to a decision-making entity.
3. The computer implementation method according to claim 2, further comprising updating a medical record with candidate text identified as AI-generated.
4. The computer implementation method according to claim 1, further comprising predicting the detection result by obtaining a detection flag indicating whether the candidate text was AI-generated.
5. Predicting the detection results is [Math 1] This further includes calculating the detection score using, where grams(Y, n) represents the set of all n grams in the sequence Y, and Y is the sequence Y of residual text. 0 and the array Y of the regenerated model text k |Y|| k | is grams (Y 0 The computer implementation method according to claim 1, wherein the normalized length of the array Yk used to normalize n) is the normalized length of n).
6. The computer implementation method according to claim 1, further comprising predicting the detection result by comparing the regenerated model text and the residual text with the model output probabilities obtained from the AI text generation model.
7. Comparing the aforementioned model output probabilities is, [Math 2] This further includes calculating the detection score using p(Y 0 |X) is the residual text array Y 0 The model output probability is given by the prefix text X, and p(Y k |X) is the regenerated model text array Y k The computer implementation method according to claim 6, wherein the model output probability is the same as the prefix text X, k is a numerical value within the sample size K, and N is the number of elements in the array.
8. Outputting explanatory text is possible with the evidence module. [Math 3] By using, it further includes calculating an evidence result (ε) between the regenerated model text and the detection result, where k is a numerical value within the sample size K, and grams(Y k , n) is the array Y of the regenerated model text k , and grams(Y 0 , n) is the array Y of the residual text 0 , and U is a union operator. The computer-implemented method according to claim 1.
9. A system for detecting AI-generated text, Memory (592) and, By using the truncation module (520), the candidate text (501) is truncated to obtain the prefix text (531) and the remaining text (531). By using the prefix text (531) with the AI text generation model (521), a regenerated model text (522) is obtained. The detection result (526) is predicted by comparing the n-gram similarity (504) between the regenerated model text (522) and the residual text (532) using the detection module (524). A system comprising: one or more processors (594) communicating with the memory (592), configured to identify whether the candidate text (501) was AI-generated by providing explanatory text (537) based on the detection result (526).
10. The system according to claim 9, wherein truncating the prefix text further comprises determining a truncation ratio.
11. The system according to claim 9, further comprising labeling candidate text in a medical record, which is identified as AI-generated, for presentation to a decision-making entity.
12. The system according to claim 11, comprising updating a medical record with candidate text identified as AI-generated.
13. Predicting the detection results is [Math 4] This further includes calculating the detection score using, where grams(Y, n) represents the set of all n grams in the sequence Y, and Y is the sequence Y of residual text. 0 and the array Y of the regenerated model text k |Y|| k | is grams (Y 0 The system according to claim 9, wherein the normalized length of the sequence Yk is used to normalize n).
14. The system according to claim 9, wherein predicting the detection result further includes comparing the regenerated model text and the residual text with the model output probabilities obtained from the AI text generation model.
15. Comparing the aforementioned model output probabilities is, [Math 5] This further includes calculating the detection score using p(Y 0 |X) is the residual text array Y 0 The model output probability is given by the prefix text X, and p(Y k |X) is the regenerated model text array Y k The system according to claim 14, wherein the model output probability is the prefix text X, k is a number within the sample size K, and N is the number of elements in the array.
16. Outputting explanatory text is possible with the evidence module. [Math 6] This further includes calculating the evidence result (ε) between the regenerated model text and the detection result by using gram(Y), where k is a number within the sample size K. k ,n) is the array Y of the regenerated model text. k And, grams(Y 0 ,n) is the array Y of the residual text. 0 The system according to claim 9, wherein U is the union operator.
17. A non-temporary computer program product comprising a computer-readable storage medium containing program code for detecting artificial intelligence (AI) generated text, wherein the program code, when executed on a computer, is transmitted to the computer. The candidate text is truncated (110), and the prefix text and residual text are obtained by using the truncation module. Using the prefix text with an AI text generation model, a regenerated model text is obtained (120), The detection result is predicted by comparing the n-gram similarity between the regenerated model text and the residual text (130), By providing explanatory text based on the detection results, it is possible to identify whether the candidate text was generated by AI (140), A non-temporary computer program product that executes a command.
18. Predicting the detection results is [Number 7] This further includes calculating the detection score using, where grams(Y, n) represents the set of all n grams in the sequence Y, and Y is the sequence Y of residual text. 0 and the array Y of the regenerated model text k |Y|| k | is grams (Y 0 A non-temporary computer program product according to claim 17, wherein the normalized length of array Yk is used to normalize n).
19. Predicting the detection result involves using the model output probability obtained from the AI text generation model for the regenerated model text and the residual text. [Number 8] This further includes comparing by calculating the detection score using p(Y 0 |X) is the residual text array Y 0 The model output probability is given by the prefix text X, and p(Y k |X) is the regenerated model text array Y k The non-temporary computer program product according to claim 17, wherein the model output probability is the prefix text X, k is a number within the sample size K, and N is the number of elements in the array.
20. Outputting explanatory text is possible with the evidence module. [Number 9] This further includes calculating the evidence result (ε) between the regenerated model text and the detection result by using gram(Y), where k is a number within the sample size K. k ,n) is the array Y of the regenerated model text. k And, grams(Y 0 ,n) is the array Y of the residual text. 0 The non-temporary computer program product according to claim 17, wherein U is the union operator.