AI Use and Limitations Disclosure

Honest framing about what TrialScope's AI capabilities are, what they aren't, and what you need to know before using AI-generated audit output in your professional work.

1. How TrialScope uses AI

TrialScope uses a foundation large language model — specifically, Anthropic's Claude — to generate audit-style analyses of clinical research documents. The AI is the primary engine of audit content generation.

When you submit a document, the platform sends the content — along with a structured prompt tailored to your selected persona, document type, review perspective, and output depth — to Anthropic's API. The AI generates structured findings. Those findings return to the platform, are rendered for you, and are stored in your account history.

The platform's value beyond the raw AI comes from the calibration layer — the prompts, document type taxonomies, regulatory framework references, persona-specific reasoning patterns, and accumulated corrections from professional users. The calibration layer steers the AI toward higher-quality, role-relevant output. It does not replace the AI; it tunes it.

2. What "persona-calibrated" does and does not mean

We position TrialScope as "persona-calibrated clinical evidence AI." Here is precisely what that means and does not mean:

It means: the prompt is structured around your selected persona, document type, perspective, and depth. The platform incorporates prior corrections from users in similar professional roles. The regulatory framework references are tailored to the document type. The output structure, tone, and emphasis are calibrated to professional conventions for your context.

It does NOT mean:

• The AI has been trained from scratch on a custom dataset. It is the same Anthropic Claude foundation model that other applications use.
• The AI is "validated" in a regulatory or clinical sense. No AI system is validated for clinical evidence review at this time.
• A real human expert has reviewed your specific audit output. The calibration layer is structural; it does not insert human review per audit.
• The platform is appropriate for regulatory submission or publication without independent human review.

We use "persona-calibrated" because it is the most accurate term we have found for what the platform actually does. We do not use it to imply more than it means.

3. What the AI does well

• Pattern recognition across large volumes of text — surfacing issues that match patterns it has seen in clinical research literature, guidance documents, and prior corrected outputs.
• Structured analysis following the conventions of the document type you're reviewing.
• Reasoning across context — connecting information from one section of a document to implications elsewhere.
• Surfacing regulatory considerations — noting which guidelines apply and where the document touches their requirements.
• Generating findings text at the depth and tone appropriate to your selected output depth.

4. What the AI does poorly or not at all

• Definitive judgment calls. The AI presents findings as professional analysis would; a qualified human must weigh those findings against the specific clinical, regulatory, and commercial context.
• Novel methodological reasoning. The AI works from patterns; first-of-kind methodological assessments require human expert judgment.
• Numerical verification. The AI may surface concerns about a calculation but should not be relied on to recompute or verify numerical results.
• Confidentiality determinations. The AI does not know what is or is not under confidentiality obligation in your specific context.
• Strategic recommendations. The AI produces an audit, not a strategy. What you do with the audit is your decision.
• Real-time regulatory updates. The AI's knowledge has a cutoff; very recent regulatory developments may not be reflected.

The calibration architecture mitigates these limitations over time as more professional corrections accumulate. It does not eliminate them.

5. Known failure modes

Large language models occasionally:

• Confabulate — generate confident-sounding output that is incorrect or fabricated. This risk is highest for specific numerical claims, citations of specific guidance section numbers, and references to specific cases or decisions. Always verify specific cited facts before relying on them.
• Miss context — overlook information present elsewhere in a document that would change the analysis.
• Apply inappropriate standards — apply standards from one therapeutic area or document type to another inappropriately.
• Reflect training biases — favor patterns common in their training data, which may not match best current practice.

The persona-calibration architecture is specifically designed to mitigate these failure modes over time by learning from professional corrections. The corrections you contribute help.

6. Your responsibilities when using AI-generated output

1. Review every finding before relying on it. Treat the audit as a sophisticated first pass, not a final product.
2. Verify specific claims independently. If a finding cites a specific section number, statistic, or regulatory requirement, verify it before acting on it.
3. Apply your professional judgment. The persona-calibration feature tunes output to your role; it does not replace your role.
4. Disclose AI use where required. Depending on your context (journal submission, regulatory filing, employer policy), you may be required to disclose AI involvement in your analysis. We do not make those disclosures for you.
5. Do not submit content you don't have the right to submit. See the Terms of Service and AUP for details.

7. Data flow to the AI provider

When you submit content for audit:

• The document content is sent to Anthropic's API over an encrypted connection
• Anthropic processes the content to generate audit output
• Anthropic returns the output to TrialScope; TrialScope returns it to you
• Per Anthropic's current API terms, content submitted via API is not used to train Anthropic's models
• Anthropic retains API request data for a limited period for abuse monitoring and operational purposes, after which it is deleted

See Anthropic's policies at anthropic.com/legal for current details.

The document content itself is not stored long-term by TrialScope. We store the audit findings (the AI's output) in your account history. We do not store the original input document content beyond the active processing session.

8. The calibration layer

TrialScope's distinguishing feature is calibration — the AI's reasoning is tuned by professional corrections made by users in specific clinical research roles.

When you correct an AI-generated finding (changing a severity, editing a finding, adding a missed finding, removing a false positive) and contribute that correction under the consent scope you select, your correction — tagged with your professional persona but otherwise anonymized — becomes part of the model's reasoning over time.

You control whether your corrections contribute via the per-contribution consent scope. Calibration contributions are anonymized — no personally identifying information about you is attached to corrections in the calibration dataset. They are tagged with your persona and verification level, and with the document type and review perspective context in which they were made.

You can opt out at any time. Existing contributions can be removed from the calibration corpus on request.

9. Questions and feedback

Questions about how AI is used in TrialScope, concerns about specific output, or feedback that should shape this disclosure: robert@trialscope.ai.

This disclosure is intentionally specific rather than aspirational. If we update it materially, we'll explain what changed and why.