Blog

The Role of AI in Supplement Recommendations

Unfair Team • January 5, 2026

AI in supplement recommendations is a decision-support tool. It filters a large evidence base through your personal context (goals, history, medications, tolerance) to surface a short list of candidates worth testing. It is not a doctor. It is not infallible. And it is only as useful as the data you give it and the critical thinking you apply to its output.

Understanding what AI recommendation engines actually do, and where they break, helps you use them well and avoid the two common failure modes: blind trust and blanket dismissal.

What an AI recommendation engine does

At its core, a supplement recommendation engine performs three operations:

1. Evidence filtering

The system maintains a knowledge base of supplement evidence: clinical trials, meta-analyses, safety guidelines, dose ranges, and known interactions. When you state a goal ("improve sleep onset"), the engine filters this base to find compounds with relevant human evidence. ^{1 2 3}

This is faster than doing the research yourself, but it is not better than the underlying data. If the knowledge base is outdated, incomplete, or sourced from low-quality studies, the recommendations will reflect those limitations.

2. Personal context matching

The engine applies your stated constraints: medications you take, sensitivities you have reported, your schedule, and your history of supplement responses. This narrows the candidate list further.

For example, if you take an SSRI, the engine should exclude 5-HTP and St. John's wort from the recommendations due to serotonin syndrome risk. ⁴ If you logged anxiety from caffeine in a previous trial, the engine should deprioritize caffeine or cap the dose.

This step is only as good as the information you provide. If you did not disclose a medication, the engine cannot screen for the interaction.

3. Feedback integration

As you log responses to recommended supplements (structured labels, side effects, adherence data), the engine updates its model of your individual response profile. Over multiple trials, recommendations become more specific to you.

This is the most valuable feature of an AI system compared to static advice. A textbook can tell you that ashwagandha may help with stress. An AI system that has your 6-week trial data can tell you that ashwagandha at 300mg produced "Stable mood" labels on 80% of your logged days, and that increasing to 600mg produced no additional benefit but caused mild GI discomfort.

Where AI recommendation engines fail

They hallucinate plausibility

AI models can generate recommendations that sound evidence-based but are not grounded in actual trial data. A recommendation for "phosphatidylserine 300mg for cortisol reduction" sounds specific and scientific. Whether it is supported by strong human evidence is a separate question. ⁵

Your defense: Check the rationale. Does the recommendation cite specific evidence? Can you verify it? If the system provides no references, treat the recommendation as unverified.

They cannot observe what you do not log

An AI system does not know that you slept terribly last night unless you tell it. It does not know that you started a new exercise program. It does not know that you are stressed about a work deadline. These confounders shape your supplement response, but they are invisible to the system unless you log them.

Your defense: Log context alongside your supplement data. Even a brief daily note ("travel day," "poor sleep," "high stress") gives the system information it needs to avoid misattributing outcomes.

They optimize for logged outcomes, not unmeasured risks

If your sleep labels improve on melatonin but you are slowly escalating the dose because the initial effect wore off, the AI may see "improving sleep" and continue recommending the approach. It does not inherently flag dose escalation as a concern unless the system is specifically designed to do so.

Your defense: Set dose ceilings and escalation rules before starting any supplement trial. Review your dose trajectory periodically, not just your outcome labels.

They inherit the biases of their training data

If the evidence base over-represents studies on young, healthy men (which much of the sports nutrition literature does), recommendations may not account for differences in women, older adults, or people with chronic conditions. If the system draws from industry-funded studies without weighting for independence, positive results may be overrepresented.

Your defense: When a recommendation matters (you have a health condition, you take medications, you are in a population underrepresented in research), verify the evidence independently. The NIH Office of Dietary Supplements fact sheets are a good starting point. ^{1 2 3}

What a good AI recommendation looks like

A recommendation worth acting on has these properties:

• Named supplement with a specific dose and timing. Not "consider magnesium" but "magnesium glycinate 300mg, 30-60 minutes before bed."
• Stated rationale linking to your goal. "Your stated goal is improved sleep onset. Magnesium glycinate has evidence for supporting sleep quality, particularly when dietary intake is low."
• Acknowledgment of your personal context. "No interactions flagged with your current medication list. You have not previously tried this supplement."
• Defined trial parameters. "Suggested trial: 21 days. Primary metric: sleep onset latency. Review at day 21."
• Explicit safety notes. "Start at 200mg for the first week. If GI discomfort occurs, reduce dose or try a different magnesium form."

If any of these elements are missing, the recommendation is incomplete and should be questioned.

What a bad AI recommendation looks like

• Vague suggestions with no dose or timing ("try ashwagandha for stress").
• Recommendations that ignore stated medications or health conditions.
• Suggestions for multiple new supplements simultaneously with no sequencing guidance.
• Recommendations with no rationale or with rationale that amounts to "this is popular."
• Any recommendation that implies a supplement can treat a medical condition.

The human role in the loop

AI narrows the search space. You make the decision. The most effective workflow is:

1. AI surfaces candidates based on your goal and context.
2. You review the rationale and check for anything the AI might have missed (interactions, personal history, practical fit).
3. You decide whether to proceed and define trial parameters if you do.
4. You log responses consistently using structured labels.
5. AI incorporates your feedback and adjusts future recommendations.

This loop gets more valuable over time. After 6-12 months of consistent use, the system has a meaningful dataset about your individual responses, and its recommendations reflect patterns that would be difficult to track manually across dozens of trials.

AI recommendations in Unfair

Unfair's recommendation engine follows this structure: evidence-filtered candidates, personal context matching (including medication screening), and continuous feedback integration from your logged responses. Every recommendation includes a stated rationale with the evidence it drew from. Recommendations you rejected or that produced adverse effects are deprioritized in future cycles. The system is designed as a decision-support tool, not a replacement for your judgment.

Continue with How AI Personalizes Supplement Recommendations, Evaluating AI Supplement Recommendations, and AI-Assisted Dose Logging.

References