Here's the tension you should feel walking into any AI companion product. Pew Research 2024 estimates more than 100 million people globally use companion AI. The 2024 Nature study of 1,006 Replika users found sixty-three percent reported reduced loneliness and three percent credited the app with preventing a suicide attempt. And the 2025 JMIR meta-analysis pooled sixty-four CBT chatbot studies showing real clinical effects. These tools work. That also means you should be picky about which ones you trust with your emotional life. This is a twelve-question consumer protection checklist. Ask every one of these before signing up, and bail if the answers don't satisfy you.

Where Do These Numbers Come From?

Sources: Harvard De Freitas 2024 study on AI companions, the 2024 Nature Replika study, the 2025 JMIR sixty-four-study CBT chatbot meta-analysis, the Stanford HAI Noora research, the MIT Media Lab 14,000-person RCT, Pew Research 2024, and the US Surgeon General's 2023 Advisory on social connection as a public health priority.

1. Does This AI Have Published Clinical Research Behind It?

Woebot has an RCT (twenty-two percent depression reduction in two weeks, JMIR Mental Health). Replika has a 2024 Nature study of 1,006 users (sixty-three percent reduced loneliness). ElliQ has published user data (ninety-five percent reduced loneliness in seniors). Ask the provider for a specific, peer-reviewed citation. If they can't name one, that's a red flag.

2. What Percentage of Users Report Positive Outcomes?

For Replika, sixty-three percent in Nature 2024. For ElliQ, ninety-five percent. For Woebot, twenty-two percent symptom reduction. Every reputable provider should be able to give you a named percentage from a named study.

3. What Does the Privacy Policy Actually Say About Chat Data?

If the provider uses your intimate conversations to train models or sells data to advertisers, that's a consumer protection issue. Read the policy. If it's vague, assume the worst.

4. Is There a Suicide Prevention Protocol?

The 2024 Nature Replika study reported three percent of users credited the app with preventing a suicide attempt, so the stakes are real. A responsible provider has crisis detection, hotline referral, and escalation. Ask explicitly.

5. Does It Use an Evidence-Based Framework Like CBT?

The 2025 JMIR meta-analysis documented sixty-four CBT chatbot studies with consistent clinical signal. CBT-trained AI has the strongest research base. If the product is just 'vibes and roleplay,' treat it as entertainment, not mental health.

6. Who Built It and What's Their Background?

Stanford HAI built Noora (thirty-eight percent empathic improvement general, seventy-one percent autistic). Dartmouth built the first NEJM-affiliated chatbot RCT. Woebot came out of Stanford clinical psychology. Look at the team credentials. Academic pedigree isn't everything, but it's a signal.

7. What Does the User Study Design Look Like?

The MIT Media Lab's AI companion trial had 14,000 participants. The Dartmouth trial was a proper RCT. The Harvard De Freitas 2024 paper had controlled comparisons. Bigger, randomized, controlled, peer-reviewed. Those are the design quality markers.

8. Does It Claim to Replace Therapy or Augment It?

A responsible product says 'augment.' The Harvard De Freitas 2024 paper makes clear AI companions supplement rather than replace human connection. If the marketing implies it can replace a psychiatrist, walk away.

9. Can You Export or Delete Your Data?

Consumer protection 101. If you can't delete your chat history or export it, you don't control the relationship. The provider does.

10. How Do They Handle Minors?

Pew Research 2024 reported two-thirds of US teens use AI chatbots. A responsible provider has age verification, minor-specific safeguards, and parental controls. The absence of these is disqualifying.

11. What's the Business Model?

Subscription and paid tiers are transparent. 'Free plus data extraction' isn't. Holt-Lunstad's body of research on social connection as a public health priority, reinforced by the US Surgeon General's 2023 Advisory, means your emotional data has real value. Know who's profiting from it.

12. Does the Provider Acknowledge Limitations Publicly?

The 2024 Harvard De Freitas paper, the Nature Replika study, and the MIT Media Lab trial all publish their caveats openly. A provider that claims no downside is either lying or hasn't done the research.

What Do You Do With This Checklist?

Go through all twelve questions before you commit. The research base on AI companions is real and growing, with sixty-four JMIR-reviewed CBT studies, a 14,000-person MIT trial, a 1,006-user Nature paper, and consistent findings in Harvard and Stanford research. The tools work. But 'works' and 'good for you specifically' are different claims, and the twelve questions above are how you bridge that gap. The fifty-seven percent of Americans Cigna says are chronically lonely deserve better than guessing.