METHOD · MODELS · LIMITATIONS
How Another AI Lie works
A plain-language account of what the checker examines, which automated systems it uses, where its evidence comes from and what its results cannot establish.
Method reviewed: 5 September 2026
The short version
Another AI Lie combines limited browser-based observations with model-assisted claim analysis and focused public-web research. It separates checkable facts from opinion or prediction, attaches accessible sources where possible, explains missing evidence and gives practical checks. It does not turn those signals into proof of truth, authorship, authenticity, fraud or plagiarism.
What happens during a free text check
1. Browser observations
The page counts limited language features such as urgency, vague authority, unsupported certainty and internal repetition. These checks happen in the browser and do not verify facts.
2. Claim separation
The evidence service identifies factual claims separately from opinions, predictions, satire and material that needs more context.
3. Public-web research
For checkable factual claims, the service can make up to four focused web-search calls and prefers primary or authoritative sources.
4. Validation and display
Only public source links returned by the search tool can be attached. Unsupported conclusions are downgraded, coverage is recalculated and concrete follow-up checks are required.
Models and services currently used
As of the review date above, the quick evidence check and Full Evidence Report use OpenAI’s gpt-5.6-luna model through the OpenAI Responses API. When factual text or a video transcript can be investigated, the request may use OpenAI’s web-search capability to locate publicly accessible pages.
Model names, service configuration and search behaviour may change as the product is tested and improved. Material changes will be recorded on this page. An AI response is never treated as a source by itself.
Where evidence comes from
There is no claim that Another AI Lie searches a complete or proprietary truth database. Evidence checks search the accessible public web at the time of the request. The system is instructed to prefer original records, official organisations and sources with direct expertise, to compare dates and context, and to look for evidence that could disconfirm a claim.
- Primary sources can include official records, government pages, original documents and statements from the organisation directly involved.
- Authoritative secondary sources can add context or independent reporting where a primary record is unavailable or incomplete.
- Repeated copies of one report are not counted as independent corroboration.
- Private databases, closed institutional systems, paywalled material, unindexed pages and deleted content may be absent.
Evidence states and coverage
| Result | Plain-language meaning |
|---|---|
| Supported | Attached evidence directly supports the claim as written. |
| Contradicted | Attached evidence directly conflicts with the claim as written. |
| Mixed | Material parts differ, or a submission contains supported, contradicted and unresolved components. |
| Unresolved | The search could not establish support or contradiction. |
| Insufficient evidence | The claim is too vague, weakly sourced or missing information needed for a responsible conclusion. |
| Needs context / Not assessed | The material is opinion, prediction, satire, quotation, non-factual or otherwise unsuitable for a factual verdict. |
Coverage describes how much of the submitted material the available evidence could address. Sufficient, partial, insufficient and not checked are deliberately separate from the verdict.
AI-writing signals
The free checker does not produce a hidden probability that text was written by AI. For predominantly Latin-script text of at least 80 words across four complete sentences, it can describe limited style observations: unusually uniform sentence length, repeated sentence openings, several formulaic transition phrases and relatively repetitive word choice in longer samples.
These patterns occur in both human and AI-assisted writing. They cannot identify an author, model, editing history or acceptable use. Short or unsupported-language samples are marked not assessed rather than forced into a score.
Similarity and repetition
Internal repetition measures repeated wording only inside the submitted text. It is not a plagiarism search. Public-source similarity uses focused searches for distinctive phrases where the material is long enough, and may label overlap exact, near, possible or common phrasing. Public overlap does not prove copying, originality, copyright infringement or misconduct.
Images, video and Full Evidence Reports
Image and video analysis is constrained by the supplied file and available context. A video report can use its transcript and representative sampled frames; it is not frame-by-frame forensic analysis. The paid report organises its analysis into overall assessment, source evidence, claim support, public-source similarity, risk signals, uncertainty and priority verification steps.
The Full Evidence Report offers depth and organisation, not certainty. See the representative sample before purchasing.
Important limitations
- Search results can be incomplete, outdated, unavailable or skewed towards well-indexed sources.
- A genuine source can still be mistaken, partial or quoted out of context.
- Absence of evidence is not evidence that a claim is false.
- A detector signal cannot establish human or AI authorship.
- Public-source similarity is not a comprehensive plagiarism database.
- High-stakes medical, legal, financial, academic, employment or security decisions require appropriate human review and specialist evidence.
Privacy, review and corrections
The checker’s handling of submitted material is described in the Privacy Policy. The publication and correction process is described in the Editorial Standards. To report a methodological error or unclear disclosure, email support@anotherailie.com.
Evidence before belief: the system is designed to show what supports a conclusion, what limits it and what should be checked next.