10 questions before you trust AI with your inner life

An AI self-development tool should not earn trust merely by producing a paragraph that feels accurate. Fluency is not evidence, empathy-shaped language is not a therapeutic relationship, and a plausible explanation is not the same as a verified one.

The American Psychological Association has urged caution around generative-AI chatbots and wellness applications used for mental health, including concerns about evidence, privacy, bias, overreliance, and unreliable crisis handling.¹ The World Health Organization has similarly emphasised autonomy, safety, transparency, accountability, and privacy in AI for health.² A non-clinical self-development product is not the same as health care, but these principles are useful wherever people may disclose intimate information or interpret outputs psychologically.

Why a buyer’s checklist matters

Many tools sit in an ambiguous category. They say they are not therapy while using clinical-sounding language. They may call an output an “assessment,” “deep psychological profile,” or “healing plan” without showing what was measured, how claims trace to inputs, or whether a qualified person has evaluated the intended use.

The checklist below does not certify a product. It helps you ask whether its boundaries are visible enough to make an informed choice.

The ten questions

1. What is the tool explicitly for—and what is it not for?

Look for a specific intended use. “Self-reflection,” “journaling support,” and “educational planning” are different from diagnosis, therapy, clinical assessment, or crisis support. A disclaimer in the footer is not enough if the sales page repeatedly implies treatment, hidden-root-cause discovery, or professional equivalence.

Ask: What should I never use this output to decide? Who should not use the tool? What happens if the input indicates immediate risk?

2. What evidence does the output use?

A generated profile may rely on a free-text prompt, a short quiz, conversation history, scored items, structured qualitative answers, or imported data. You should know which. Specificity can come from your own detailed input without proving that the interpretation is valid.

Prefer outputs that distinguish quoted or paraphrased evidence from inference. If a claim cannot be traced to something you provided or a clearly described rule, treat it cautiously.

3. Does the system separate observations from interpretations?

“You described delaying three conversations after receiving critical feedback” is an observation grounded in self-report. “You avoid conflict because rejection felt unsafe in childhood” is an interpretation containing a cause. The second may feel coherent and still be unsupported.

A responsible tool should mark uncertainty, preserve plausible alternatives, and avoid filling gaps with detail that merely sounds psychologically literate.

4. Can you disagree without being pathologised?

Be wary of systems that turn disagreement into confirmation: “Your resistance proves the pattern,” “You are not ready to see the truth,” or “The result knows you better than you know yourself.” This creates an unfalsifiable loop.

You should be able to reject, refine, or correct an interpretation. The product should explain what happens when you report a factual error or an unsupported claim.

5. What personal data is collected, retained, and shared?

Read the privacy policy before entering sensitive information. Look for the specific categories collected, service providers involved, storage period or deletion logic, access controls, international transfers, and how to request access or deletion.

The US Federal Trade Commission has warned AI companies to honour privacy and confidentiality commitments, especially where customers may provide sensitive or confidential information.³ A vague assurance that data is “secure” does not answer who receives it or why.

6. Are your answers used for marketing or model training?

These are separate purposes from delivering the service you requested. Look for an explicit answer. If data may be used to train models, improve products, build advertising profiles, or create testimonials, ask whether consent is optional, specific, and revocable.

A feedback request is not automatic permission to publish a quote. Testimonial permission should be separate and should specify whether the quote is exact, anonymous, or named.

7. What quality checks happen before you see the result?

“AI-powered” describes a mechanism, not a quality system. Ask whether outputs are checked for unsupported certainty, contradictions, duplication, unsafe recommendations, fabricated facts, or privacy leakage. Ask what happens when a check fails.

NIST’s AI Risk Management Framework and Generative AI Profile emphasise managing risks through governance, mapping, measurement, and management rather than assuming a model is trustworthy by default.⁴ Consumer products will implement this differently, but they should be able to describe their controls in plain language.

8. Where is a human involved—and where are they not?

“Human-reviewed” can mean every output is read by a qualified professional, a sample is audited, a support agent can intervene, or only the system design once involved a person. Ask who, when, and with what credentials.

Equally, a product should not imply a human relationship that does not exist. An automated response cannot observe nonverbal information, carry professional duties merely by sounding caring, or reliably manage a crisis.

9. Can you access, correct, export, and delete your material?

A psychologically personal output should not become a locked black box. Look for export and deletion options, a correction route, retention details, and what happens to derived documents if a source record changes. If you paid, check refund and correction terms before purchase.

10. What does the business model reward?

Does the tool benefit when you keep chatting, disclose more, buy repeated readings, or become dependent on daily reassurance? Or does it have a bounded task and a clear end state?

Engagement is not always harmful, and a one-time product is not automatically safe. But incentives shape design. Notice whether the copy increases agency or cultivates fear that only the system can explain you.

Red flags that deserve a pause

• “Clinically accurate,” “validated,” or “therapist-level” without a study matching the actual product and intended use.
• Claims to identify trauma, diagnosis, subconscious truth, or childhood causes from a brief quiz or chat.
• No clear operator, privacy policy, contact route, retention terms, or deletion process.
• Testimonials used as proof of typical accuracy or therapeutic benefit without adequate context.
• A result that contains precise life details you never provided.
• Advice to stop treatment, ignore professional guidance, confront someone, end a relationship, quit work, or make a major financial move based mainly on the output.
• Prompts that solicit trauma or crisis disclosures for an entertainment-style result.
• Language that makes disagreement impossible.

How AXIS answers its own checklist

AXIS is educational self-development, not therapy, medical advice, clinical assessment, or crisis support. The free Mirror quizzes use fixed weights and local browser scoring; they do not use AI to calculate or display results. The paid Diagnostic uses AI to assist synthesis of a structured 92-question intake into a Portrait.

The system is instructed to separate answer evidence from interpretation, frame interpretations as working hypotheses, state uncertainty, include plausible alternatives, avoid diagnosis, leave unsupported gaps unfilled, and never treat disagreement as proof of resistance. The complete Diagnostic must pass automated quality and safety checks before release; flagged work is held rather than delivered unchecked.

The 90-Day Companion is rendered deterministically from the validated Portrait, and the Complete Text Edition is generated mechanically from the Portrait and Companion. Neither derivative makes a separate AI interpretation. Diagnostic answers and documents are not used for AXIS marketing or model training. The Privacy Policy, AI & Safety page, and Terms describe the current boundaries and correction process.

Those controls do not make AXIS infallible or clinically validated. The result remains a non-clinical interpretation of self-report, which you should test, reject, or refine.

If you want to see the offer after reviewing its boundaries, visit the AXIS Diagnostic. If you prefer to start without AI or payment, use the free AXIS Mirror.