AI Disclosure

1. Which AI we use

Driora calls one model: Claude, made by Anthropic. We use Claude through Anthropic's commercial API. We do not use OpenAI, Google Gemini, Meta Llama or any other model. We do not host our own model.

The specific Claude model version used for any given report is recorded in the report metadata under _meta.model so you always know exactly what produced your output.

2. What the AI sees

For each report we send Claude exactly three things:

• The job description text we extracted from the URL you provided (or that you pasted into the form).
• Your resume content, in plain text. PII like your phone number is redacted before the prompt is sent. Your name and email do remain in the context so the cover letter can be addressed to the right hiring manager.
• The role context we parsed from the URL: company name, job title slug, source platform. Nothing else from the URL beyond these fields.

Those are the inputs about you. For the company-research section the AI also runs live web searches about the target employer (the company and role, never you); those queries are screened to exclude your personal data, and the company facts in your report cite the real published sources they came from. The exact prompt template lives in the open source we are still in alpha for, so we are happy to walk you through it on request.

3. What the AI does not see

The AI never sees:

• Your email inbox or any messages.
• Other reports you ran in the past or the resumes from those reports.
• Any browsing history or analytics.
• Any payment information (we collect none).
• Other users' resumes, reports or job histories.

Each report is generated in a fresh, isolated request to Claude. There is no shared context between two of your reports and no shared context between your report and any other user's. Anthropic's API is stateless from our side: the model has no memory of prior requests.

4. Four guards against hallucination

Hallucination, where the model invents plausible-sounding facts, is the central failure mode of any AI tool that produces business intelligence. We layered four guards so it cannot happen quietly.

Guard 1: the job posting must be readable

Before we even call Claude, we verify we successfully extracted the job description text from your URL. If the URL is auth-walled, JS-rendered or otherwise unscrapeable, we return an error and tell you to paste the description text directly. We do not generate a report from a URL alone.

Guard 2: the prompt requires real content

The prompt builder refuses to construct a Claude request when the job description is empty. There is no placeholder text that says "use web search to find this posting" because that placeholder is the exact instruction that produced fabricated reports in the past.

Guard 3: the system instructions ban fabrication

Our system prompt tells Claude that the job description is the single source of truth. Claude is explicitly told not to invent funding rounds, headcount changes, partnerships or leadership facts unless the source posting states them. The system prompt also bans the speculative "given the candidate's background, this is likely for..." framing that signals the model is guessing.

Guard 4: a quality gate inspects every finished report

After Claude returns, we run a fingerprint scan on the output. If the report contains the speculative-framing pattern, an invented funding figure not present in the source job description, an invented headcount narrative or a leftover error placeholder, we block the email and surface a structured failure instead. The gate is open source defensive code; you can read the fingerprint list in our repository.

5. Your data is not used to train any model

Anthropic's commercial API does not use customer prompts or completions to train Claude or any future Claude. This is the default for all paid API access and it is the contract we operate under. See Anthropic's privacy commitments for the canonical reference.

Driora itself never trains any model on your data either. We do not build models. We do not sell training corpora. Your resume, job URL, name and email exist only to produce your report and any follow-up emails you opted into.

6. Every report is labelled as AI-generated

The body of every report carries a metadata block that includes:

• is_ai_generated: true
• provider: anthropic
• model: <model name and version>
• generated_at: <timestamp>
• human_review_available: true

Email reports include the same disclosure in the footer. We do not strip these labels under any circumstance. If you forward a Driora report to a recruiter, the AI provenance travels with it.

7. Human review and corrections

If a report contains a factual error about you or about the role, email [email protected]. We read every correction request, regenerate the report when warranted and use the failure mode to tighten the four guards above. The pattern of corrections is what makes the next version better.

We do not currently offer a one-click "regenerate this report" button. That is on the alpha roadmap.

8. Questions or concerns

Reach Driora directly: [email protected]. For privacy-specific questions see the Privacy Policy. For how we differ from competitors who use AI to mass-apply on your behalf see the Privacy Comparison.