Why Do AI Answers Miss Harms and Side Effects?

In the last few years, we've seen a dramatic shift in how people search for health information. Traditional search engines — where you type keywords and get links to trusted sources — are increasingly being supplemented or even replaced by generative AI chatbots powered by large language models (LLMs). These tools can provide direct answers in a conversational style, seemingly making health knowledge more accessible.

But there's a catch. When it comes to the delicate balance of benefits and harms in health information, AI-generated methodological rigour responses often fall short. They may gloss over side effects or risks, omit uncertainty, or even present hallucinated facts and citations — creating serious issues of health misinformation. This blog dives into why that happens, why trust is the key limiting factor, and what reputable organizations like Cochrane and The Cochrane Collaboration contribute to this landscape.

From Search Engines to Generative AI Chatbots: What’s Changing?

Traditionally, if you had a health question, you might:

Use a search engine (Google, Bing, etc.) to find reputable articles or sites like the Cochrane Library, NHS, CDC, or peer-reviewed journals.
Click through to expert-curated evidence or systematic reviews.
Read detailed information that outlines both benefits and harms, discusses limitations, and references research quality.

Now, with generative AI chatbots powered by LLMs, the interaction looks quite different:

You ask a question in natural language.
The AI generates an answer in sentences or paragraphs, often summarizing concepts or synthesizing information.
Responses feel conversational and sometimes more accessible.

This shift can be useful but raises critical questions about trust. How much can you rely on an AI assistant that does not cite its sources transparently or clearly indicate the uncertainty around the answer?

Why Trust is the Limiting Factor in AI-Generated Health Information

In health communication, trust isn’t just about believing the information; it’s about evaluating its reliability and relevance to your situation. Reputable tools and collaborations have set a high bar for trustworthy information:

The Cochrane Collaboration is renowned for rigorous systematic reviews that summarize all available evidence, emphasizing transparency, quality, and balanced reporting of both benefits and harms.
The Cochrane Library is a gold-standard source, offering systematic reviews with detailed methods, risk of bias assessments, and clear grading of evidence strength.

In contrast, LLM-based chatbots don’t “understand” health information. They generate text probabilistically — predicting what’s likely to come next in a sentence given the prompt and their training data — but they do not verify whether the content is true or balanced.

Implications:

AI can repeat common knowledge or prevalent narratives without critical appraisal.
Risks and side effects may be underrepresented or omitted because they’re less common or less emphasized in training data.
Potential harms from medicines, procedures, or lifestyle changes can be glossed over, which poses a safety concern.

LLMs Predict Text, They Do Not Understand Truth

Large language models, including those underpinning many AI chatbots, work by learning patterns across vast datasets of human-generated text. They excel at generating coherent and contextually relevant text, yet they operate without consciousness or factual verification.

This fundamental limitation leads to several issues relevant to health queries:

Hallucinations: The AI may invent facts, statistics, or citations that seem plausible but are entirely fabricated.
Incomplete Answers: The AI may omit important details, such as side effects or contraindications, because these nuances are difficult to capture in broad datasets or were not prominent during training.
Lack of Source Transparency: AI outputs often lack clear, verifiable citations, which makes it difficult for users to check accuracy or explore further.

Example: Hallucinated Citations

One particularly concerning problem is when AI chatbots fabricate DOIs, article titles, or author names to mimic academic references. Someone seeking evidence-based reassurance may be misled into thinking the information is sourced from authoritative research when, in fact, it is not.

As someone who has double-checked dozens of DOIs and citations, I can’t stress enough that a fake reference is a red flag. AI-generated health information that invents sources contributes to health misinformation and undermines trust in science.

The Balance of Benefits and Harms Is Complex and Requires Rigorous Evidence Synthesis

Health decisions often hinge on weighing benefits versus harms:

Choosing a medication involves balancing symptom relief with potential side effects.
Screening tests have benefits in early detection but risks from false positives and subsequent interventions.
Lifestyle advice is coupled with potential unintended consequences.

The Cochrane Collaboration’s strength is in its systematic approach to evidence synthesis, which explicitly highlights that complexity. Their reviews:

Gather all relevant studies meeting predefined criteria.
Assess risk of bias in those studies.
Clearly summarize main benefits and harms, including the quality of evidence.
Provide balanced, nuanced conclusions to guide clinicians and patients.

Such rigor is almost impossible for current generative AI chatbots, which produce answers based on pattern recognition, without systematic methodology.

What Can Users Do When Facing AI-Generated Health Information?

While generative AI chatbots are powerful tools, users should remain cautious, especially where health decisions are concerned:

Look for transparency: Does the answer provide clear sources? Is it drawing from trustworthy evidence syntheses like those from the Cochrane Library?
Check for mention of harms and side effects: Balanced information always includes potential downsides, uncertainties, or limitations.
Consult multiple sources: Use AI-generated summaries as a starting point, not a conclusion. Cross-reference with established guidelines or trusted databases.
Discuss with healthcare professionals: AI tools can augment knowledge but cannot replace personalized advice from qualified clinicians.

Looking Forward: Combining AI with Trusted Evidence Sources

There is exciting potential to enhance AI by integrating it with trusted repositories like the Cochrane Go here Library. Some possibilities include:

Training AI models explicitly on systematic reviews and evidence summaries to improve accuracy.
Building interfaces where AI-generated responses are always linked to verified citations.
Developing AI tools to aid evidence synthesis rather than replace it, supporting human experts.

Until then, we must recognize the current limits of AI chatbots and guard against inadvertent spread of health misinformation, especially with respect to harms and side effects.

Summary

The rise of generative AI chatbots and large language models is transforming health information access. However, their fundamental nature — predicting text rather than understanding truth — means they often miss or minimize potential harms and side effects. Trust, built through transparent sourcing and balanced evidence synthesis such as that provided by The Cochrane Collaboration and the Cochrane Library, remains crucial when evaluating AI-generated health information.

Being aware of these limitations empowers users to critically appraise AI chatbots’ answers, seek balanced information on both benefits and harms, and avoid health misinformation pitfalls.

If you want evidence you can trust, look for resources developed with rigorous standards and expertise — the very foundation of high-quality health knowledge.