Help, Hype, and the Hard Truth About Safety
It is two in the morning.
Your thoughts are racing. Your chest feels tight. You do not want to wake anyone. You do not want to explain yourself. You open your phone and type into a chatbot:
“I cannot cope.”
Within seconds, a reply appears. Calm. Structured. Reassuring.
For many people, artificial intelligence now feels like a lifeline. It is private. It is immediate. It does not judge. It does not sigh. It does not look uncomfortable.
But here is the truth we need to hold carefully and honestly.
AI can support mental health in certain ways.
And in other ways, it can be genuinely risky.
This is not a story about fear. It is not a story about blind optimism either. It is about clarity.
What Do We Mean by AI for Mental Health?
When people talk about AI in mental health, they usually mean one of three things.
First, structured mental health apps or chatbots that deliver tools based on psychological models such as cognitive behavioural therapy. These might include mood tracking, behavioural activation prompts, coping exercises, or thought reframing techniques.
Second, general purpose conversational AI tools that people use informally for emotional support, journalling, advice, or reflection.
Third, clinician facing systems that help with administration, documentation, triage support, or screening processes.
These are very different categories. They do not carry the same risks. They do not have the same evidence base.
Professional bodies have begun raising important questions about safety and governance. The American Psychological Association has issued consumer guidance highlighting risks linked to generative AI tools used for mental health advice. The World Health Organization has published ethical guidance on artificial intelligence in healthcare, focusing on transparency, accountability, and data protection. In the United Kingdom, the National Institute for Health and Care Excellence has updated its evidence standards framework for digital health technologies to include AI driven systems.
This tells us something important. The conversation has moved from curiosity to regulation. That is a sign of maturity, but also a sign that caution is needed.
What AI Genuinely Does Well
Let us begin with what works.
1. Access and Immediacy
AI is available at any hour. There are no waiting lists. No referral processes. No appointment delays.
For someone feeling ashamed, frightened, or unsure whether their distress is “serious enough,” that immediacy can reduce the first barrier to speaking out.
Some surveys suggest that a notable minority of teenagers admit sharing things with AI that they have not told friends or family. That does not automatically mean this is healthy. But it does reveal something about perceived safety.
Naming a feeling is often the first step toward regulating it.
2. Structured Self Help
Certain AI chatbots designed around cognitive behavioural techniques have shown small to moderate improvements in mild to moderate symptoms in early research.
For example, tools developed by organisations such as Woebot Health have been evaluated in controlled trials examining reductions in depressive symptoms over short periods. Similar research has been conducted on CBT informed conversational agents developed in university settings internationally.
The effects are not dramatic. They are not a replacement for therapy. But they are not meaningless either.
When a tool reliably prompts behavioural activation, sleep hygiene routines, or structured thought reflection, it can support habit change.
3. Skills Practice and Behavioural Support
AI can be particularly useful for:
- Journalling prompts
- Grounding exercises
- Breaking tasks into steps
- Drafting difficult conversation scripts
- Creating coping cards
- Encouraging daily routines
- Tracking sleep and mood
In many ways, this is similar to having a structured workbook that talks back.
For clients already in therapy, AI can help organise reflections between sessions. It can assist with identifying triggers or patterns. Used this way, it becomes supplementary.
4. Lowering the Threshold for Help Seeking
Some individuals will not attend a GP appointment. They will not call a helpline. They will not step into a therapy room.
They may, however, open an app.
From a public health perspective, that matters.
Where Things Become Risky
Now we need to look at the other side.
AI systems can sound confident even when they are wrong. They do not possess clinical judgement. They do not hold ethical accountability in the way a registered professional does.
1. Crisis Situations
AI tools can struggle with nuanced risk assessment. In situations involving suicidal ideation, escalating self harm, psychosis, mania, or severe eating disorders, subtle cues matter.
A human clinician assesses tone, pacing, inconsistencies, history, safeguarding concerns, and immediate risk. An AI system analyses patterns of language.
Those are not equivalent processes.
In 2025, leaders within NHS England publicly warned young people against relying on AI chatbots as substitutes for therapy, particularly in high risk contexts.
2. Confident Misinformation
Generative systems can fabricate details. In mental health contexts this might look like:
- Suggesting a coping strategy that is inappropriate
- Labelling normal stress as a psychiatric disorder
- Giving medication advice without qualification
- Missing red flags
When information is delivered fluently, it can feel authoritative.
That is where critical thinking becomes essential.
3. Emotional Over Reliance
If AI becomes the primary attachment figure for emotional regulation, difficulties can emerge.
Constant reassurance seeking can reinforce anxiety cycles. Avoiding real world conversations can deepen isolation. Dependency can develop quietly.
Mental health recovery is relational. Human connection remains central.
4. Privacy and Data Concerns
Many users assume that typing into an AI system is private because there is no visible human. In reality, privacy depends on data policies, storage practices, and governance structures.
Mental health disclosures are deeply sensitive. They deserve careful protection.
5. Bias and Cultural Limitations
AI systems are trained on large datasets that may not represent all communities equally. This can influence tone, assumptions, and relevance of advice.
Subtle bias can shape responses in ways that are difficult to detect but impactful over time.
Practical Guardrails for Safer Use
If someone chooses to use AI as part of their mental health toolkit, some practical boundaries are helpful.
Use AI for skills, not diagnosis.
Coping strategies, journalling prompts, behavioural planning, and communication rehearsal are reasonable uses. Diagnostic interpretation and medication decisions are not.
Look for transparency.
Is there published research? Are limitations stated clearly? Is crisis signposting included?
Protect identifying information.
Avoid sharing personal data you would not send in a standard email.
Reality check extreme advice.
If a response suggests drastic actions, rigid conclusions, or absolute certainty, pause and verify with a human professional.
Think of AI as first aid, not surgery.
It can help stabilise, organise, and prompt reflection. It is not designed for trauma processing, complex safeguarding issues, or severe psychiatric conditions.
When to Go Human
There are clear circumstances where AI should not be your main support:
- Suicidal thoughts or escalating self harm
- Psychosis like experiences
- Signs of mania
- Severe eating disorder symptoms
- Feeling unsafe at home
- Worsening depression despite self help
- Medication concerns or side effects
- Reassurance seeking that feels compulsive
In these situations, human clinicians offer real time risk assessment, ethical accountability, relational depth, and coordinated care.
AI does not replace that.
A Balanced Conclusion
Artificial intelligence is not inherently dangerous. It is not inherently therapeutic either.
It is a tool.
Used carefully, it can provide structure, accessibility, and supplementary skills practice. Used without boundaries, it can increase risk, misinformation, dependency, or delay appropriate care.
The most grounded position is neither panic nor hype.
It is informed caution.
If we treat AI as we would treat a gym plan, it can support training and discipline. But when something is injured, complex, or deteriorating, we seek a qualified professional.
Mental health deserves that level of seriousness.