Large language models. The transcription and interview-analysis features use leading large language models from major AI providers (for example OpenAI, Anthropic or Google), accessed through their business services under contracts that prohibit the provider from using interview data to train or improve its models and require it to process that data only to provide the service to us. The Deepfake Check uses our own proprietary technology.

3. What the AI does not do

The following apply to every feature of the platform:

• It does not analyse facial expressions, eye movement or gaze, body language, posture or gestures.

• It does not analyse tone of voice, accent, pitch, speech rate or any other characteristic of how a Candidate sounds. Audio is used only to produce the transcript.

• It does not analyse or score a Candidate's appearance, clothing, background or surroundings.

• It does not attempt to infer or predict age, sex, gender, race, ethnicity, nationality, disability, health, religion, sexual orientation, political opinion or any other protected or sensitive characteristic, and it does not use any such characteristic in scoring.

• It does not identify Candidates or match faces against any database. The Deepfake Check looks for signs of manipulation within the video itself; it does not compare the Candidate to any other image or record.

• It does not make decisions. The platform produces information; the Customer decides.

• It does not learn from Candidates. No interview recording, transcript, result or report is used to train, retrain or fine-tune any AI model.

4. Human oversight and the Candidate's position

Customers. Each Customer agrees with us, as a condition of using the platform, that: (a) AI outputs are decision-support only; (b) no decision with legal or similarly significant effects on a Candidate is made solely on the basis of them without meaningful human review; (c) Candidates are told that AI-assisted analysis is used; and (d) outputs are not used in a way that discriminates on the basis of any protected characteristic. Every report carries a notice that it was generated with AI assistance and should not be the sole basis for a decision, and Customers must preserve that notice when sharing reports.

Deepfake Check in particular. The Deepfake Check produces a likelihood indicator, not a finding of fact. Customers must treat it as one input requiring human judgement, must not reject a Candidate solely on the basis of an integrity score without further verification, and must give the Candidate an opportunity to respond where a score influences a decision.

Candidates. Before starting an interview, Candidates are shown the Candidate Terms and Candidate Privacy Notice, which describe the AI features in plain language, and are asked for explicit consent to the Deepfake Check (and, where enabled, the personality questionnaire). A Candidate who does not consent does not take the interview on the platform and the organisation that invited them has agreed to offer another way to be considered. Candidates can ask the inviting organisation for human review of any AI-assisted assessment, express their point of view and contest a decision. Because the inviting organisation makes the decisions and controls the data, requests go to it; we cannot review assessment outcomes ourselves, but if a Candidate contacts us we will pass the request on promptly.

5. Limitations

• Transcription can contain errors, particularly with background noise, poor audio quality, strong accents, technical vocabulary or names. Scores are derived from the transcript and may therefore be affected by any error in it. Our scoring instructions are designed to direct the model to treat apparent transcription errors — for example an implausible word in an otherwise coherent answer — as such, and not to count them against the Candidate. This mitigates but does not eliminate the effect of transcription errors, which is one reason the Platform's outputs are decision-support only and a human reviewer should refer to the recording where a score appears inconsistent with the answer given. 

• Language-model output can contain errors, omissions and biases inherent in such models and in the questions and scoring criteria configured by the Customer. Scores and recommendations are not professional, legal or HR advice.

• The Deepfake Check is probabilistic. The Deepfake Check is probabilistic. It can produce false positives (flagging a genuine video) and false negatives (missing a manipulated video). Poor lighting, low-quality cameras, heavy compression, virtual backgrounds and some assistive technologies can affect the result. We do not warrant that any video is or is not authentic.

• The Big Five assessment is a self-report tool. It is not a clinical, medical or diagnostic instrument, does not measure ability or integrity, and must not be the sole basis for any decision.

• The quality of interview analysis depends on the quality and relevance of the Customer's questions and criteria. The platform does not validate whether a Customer's criteria are appropriate for a given role.

6. Testing, bias and ongoing review

Before release, each AI feature is tested against internal test sets to confirm that it performs its intended function. We have not yet carried out a formal adverse-impact or bias audit of the interview-analysis feature; we plan to do so as the platform accumulates sufficient real-world data to make such testing statistically meaningful, and we will update this statement when that work is done. Until then, the safeguards on which we rely are the design choices described in section 3 (content-only analysis; no inference of protected characteristics), the human-review requirements in section 4, and the Customer's own obligations under equality law. We monitor the performance of the Deepfake Check against newly emerging generation techniques and update it as those techniques change. Customers conducting their own equality or validation analysis can request reasonable technical information from us to support it.

7. Data protection and regulatory position

• The organisation that invites a Candidate is the controller of the Candidate's personal data; we process it as that organisation's processor. The Customer determines the lawful basis for the Deepfake Check and the personality assessment and carries out a data protection impact assessment where one is required. We provide reasonable information to support it.

• Interview data is hosted in the United Kingdom.

• The Deepfake Check may, depending on the jurisdiction and configuration, involve processing of biometric or special category data. Explicit consent is collected from the Candidate before every interview.

• AI systems used in recruitment and candidate evaluation may be classified as high-risk under the EU AI Act. The platform is not currently offered in, or directed at, the EU or EEA. Customers who use it in other jurisdictions are responsible for any local AI-in-hiring requirements (for example New York City Local Law 144 or the Illinois Artificial Intelligence Video Interview Act) and we provide reasonable technical documentation to assist them.

• We are responsible for any obligations that apply to us as a provider under applicable AI legislation and will update this statement as that law develops.
  
  

8. Questions and changes

Questions about this statement or about how AI is used in the platform: privacy@nextgeninterviews.ai. Questions from Candidates about a specific interview or decision should be directed to the organisation that invited you. We will update this statement when the AI features change materially; the version number and date at the top identify the current version. 

Precise AI Limited, Level 1, Devonshire House, 1 Mayfair Place, London, W1J 8AJ, United Kingdom. Company no. 16512338.