How Do You Know When AI Is Giving You Something You Can Trust?
In the previous blog, I explored the possibility that AI can be useful for emotional reflection — not as an authority on our lives, but as a tool that may help us pause, think, become curious and develop greater discernment.
That leaves us with another question.
When AI gives us a response that sounds insightful, how do we know how much confidence to place in it?
This is not always easy.
AI can produce language that is clear, emotionally attuned and remarkably persuasive.
Sometimes a response immediately resonates.
You may find yourself thinking:
Yes. That's exactly it.
And perhaps it is.
But perhaps it is only one possible way of understanding what you have described.
Perhaps part of it is well grounded in established knowledge. Perhaps another part is interpretation. Perhaps AI has noticed an interesting pattern worth exploring. Or perhaps the response simply sounds convincing because AI is extraordinarily good at generating coherent language.
These are not the same things, and learning to recognise the difference is an important part of using AI reflectively.
Coherence can feel like truth
Human beings are influenced by coherence.
When an explanation is clear, confident and well organised, we tend to experience it as credible. A good story feels as though the pieces fit together.
AI is particularly powerful in this respect because it happens that language models are exceptionally good at creating responses in which the pieces appear to fit.
That does not mean the information is false, but neither does coherence make it true.
AI can produce beautifully constructed language without possessing human understanding or lived experience. It can organise information persuasively while still working with incomplete information, assumptions contained within our question, or patterns in language rather than established fact.
This becomes particularly important when we use AI for emotional reflection.
Imagine telling an AI assistant:
Every time someone becomes distant from me, I panic. I think I must have an anxious attachment style.
The response may explain anxious attachment beautifully. It might describe hypervigilance to signs of rejection, sensitivity to changes in another person's availability, or difficulty tolerating relational uncertainty. You may recognise yourself strongly in the description.
But what has actually been established?
Perhaps some of the information about attachment is well supported.
Perhaps there is a reasonable possibility that attachment experience is relevant to what you described.
But the AI does not know enough about your life to determine whether this explanation is the most useful one. There may be other possibilities.
Your current relationship may genuinely be unpredictable.
A recent loss may have heightened your sensitivity.
Your nervous system may be under considerable stress.
An old relational experience may have become activated.
Or several things may be occurring at once.
A compelling explanation can be useful. It becomes less useful when possibility quietly turns into certainty.
Discernment is not the same as doubt
The answer is not to distrust everything AI says. That would simply move us from one extreme to another.
Discernment is different. Discernment asks us to remain open while also remaining thoughtful.
We can receive something useful without immediately adopting it as truth. We can notice that an idea resonates without needing to build an identity around it. We can hold an interpretation lightly enough to ask:
What supports this?
What else might be possible?
What doesn't quite fit?
What information might be missing?
How does this sit alongside my actual lived experience?
This capacity may not come easily to everyone.
Some of us have become accustomed to looking outside ourselves for certainty.
Others have learned to mistrust their own perceptions.
And when we are emotionally distressed, frightened, lonely or confused, a clear explanation can feel especially relieving. That relief is real.
It is also one reason it can be helpful to have a way of slowing the interaction down.
How to assess the reliability of AI responses
While developing Using AI Reflectively and Safely, I wanted a simple way of distinguishing between different kinds of information and reflection that can appear in an AI conversation.
The result was an AI Reliability Framework for reflective and relational enquiries.
It uses five markers:
GK — Grounded Knowledge
Information that is well established and supported by strong evidence or broad agreement.
If AI is making this kind of claim, asking for information sources can help us examine where the knowledge comes from.
II — Informed Inference
A reasoned interpretation based on existing knowledge, experience or therapeutic understanding.
An informed inference may be thoughtful and useful.
But an inference remains an inference.
It deserves to be held differently from established knowledge.
EP — Exploratory Pattern Recognition
An interesting reflection, theme or conceptual connection that may invite curiosity and exploration but is not necessarily established truth.
This is often where AI can be particularly interesting reflectively.
It may notice a connection that helps us wonder about something differently.
The important word is exploratory.
A pattern is something to become curious about, not something we automatically need to believe.
LC — Language Completion
A response that sounds coherent and convincing because the language flows well, even though the underlying information or reasoning may be limited.
This is an especially important one.
AI is designed to generate coherent language.
Sometimes fluency itself creates an impression of knowledge or authority that the underlying response does not deserve.
UF — User Framing
A response that has been influenced by the way we framed our question, the emotional tone we brought into the interaction, or the narratives already present in the conversation.
This brings us back to something we touched on in the previous blog.
If I ask:
Why does my partner keep manipulating me?
the word manipulating has already shaped the enquiry.
AI may begin reasoning from a conclusion I supplied.
If instead I say:
Something happened with my partner that I am interpreting as manipulation. Can you help me explore what I noticed and consider several possible explanations? I have created much more room for reflection.
The framework isn't about getting the “right” answer
This distinction matters to me.
The purpose of a reliability framework is not to make every AI conversation into an intellectual fact-checking exercise. Nor is it designed to remove emotion from reflection.
If you are talking about something painful, confusing or deeply personal, you are still allowed to be affected by it.
The purpose is to create a little structure around the conversation so that emotion can remain present without requiring us to surrender discernment.
We can feel:
That response really touched me.
And also wonder:
What kind of response was it?
We can recognise:
That explanation makes an enormous amount of sense to me.
And still ask:
How much confidence does it deserve?
Both can exist together.
In fact, that ability to hold emotional experience alongside curiosity may be more valuable and a life skill than achieving certainty.
You can ask AI to help you discern its own responses
There is a pleasantly paradoxical possibility here.
We can ask AI to help make the limitations of its responses more visible.
For example, before beginning a reflective conversation, you might give your AI assistant the five Reliability Framework categories and ask it to identify the nature of the material it is offering as the conversation unfolds.
You might ask:
Which parts of this response are grounded knowledge?
Where are you making an informed inference?
Is there anything here that is better understood as exploratory pattern recognition?
Could any part of this sound more certain because the language is coherent than the evidence warrants?
How has the way I framed my question influenced your response?
These questions introduce a useful interruption into the conversation.
Instead of simply moving from one persuasive response to the next, we create a small pause.
And in that pause, we have another opportunity to reflect.
Whilst there is only anecdotal evidence for this framework, the experience of working with it seems to indicate a potential to remain more grounded in the present, than without it.
Our framing matters more than we may realise
The User Framing marker is especially important in emotional conversations.
We never arrive at an experience entirely neutrally. We bring our fears, our hopes, our assumptions, attachment experiences, emotional states and existing interpretations. That is simply part of being human.
If I am frightened of rejection, I may naturally describe an interaction through that lens.
If I already believe someone doesn't care about me, I may present evidence that supports that belief while leaving out details that complicate it.
If I feel ashamed, I may ask questions that begin from the assumption that something is wrong with me.
AI only knows what we place into the conversation. It cannot stand in the room where the interaction occurred. It cannot hear the other person's account. It cannot know everything we have forgotten, misunderstood, omitted or not yet recognised ourselves.
So sometimes one of the most useful questions we can ask is:
“What assumptions are already contained in the way I have asked this?”
That question is valuable well beyond AI.
It is a reflective question about how all of us make meaning.
Sometimes not knowing is the most accurate position
There is a strong human pull towards explanation.
Something happens. We feel unsettled. We want to understand - uncertainty can be uncomfortable.
AI can reduce that discomfort very quickly because it can nearly always produce an explanation.
But sometimes the most grounded response to a human experience is:
We don't know yet.
Perhaps we need more information. Perhaps we need to notice what happens next. Perhaps there are several plausible explanations. Perhaps our understanding will change as our emotional intensity settles. Perhaps something needs to be explored in therapy or in conversation with the person involved. Or perhaps meaning will emerge gradually through experience rather than through analysis.
Discernment includes the capacity to leave some questions open.
It allows us to say:
This is interesting.
This may be relevant.
I want to explore this further.
I'm not ready to decide what it means.
That is not a failure to understand. Sometimes it is a very sophisticated form of understanding.
Developing discernment rather than borrowing certainty
This is why I think the larger value of the Reliability Framework lies beyond deciding whether AI is “right”.
It gives us an opportunity to practise discernment.
We begin noticing the difference between knowledge and interpretation. Between an interesting pattern and an established conclusion. Between something that feels emotionally resonant and something we have reason to regard as reliable. Between a response that expands our awareness and one that simply reflects our existing assumptions back to us.
Over time, this practice can strengthen something important. That is, our capacity to remain open to what is offered while continuing to reference evidence, context, lived experience and our own developing understanding.
AI does not need to become the place where certainty lives.
It can instead become one of the places where we practise asking better questions.
Perhaps the question is therefore not simply:
“Can I trust this AI response?”
A more useful question may be:
“What kind of response am I receiving, and how much confidence does it deserve?”
That leaves room for knowledge. It leaves room for interpretation, it leaves room for curiosity. And most importantly, it leaves room for us.
And where does therapy fit?
Once we begin making these distinctions, another question naturally emerges.
If AI can provide information, reflection, questions, emotional language and sometimes genuinely helpful perspectives, what remains particular to psychotherapy and human therapeutic relationships?
The answer is much more substantial than simply saying that a therapist is a better source of information. Therapy offers something AI cannot reproduce through language alone.
That is where we'll go next.
A note about this reflection
This material is offered for educational and reflective purposes only. It is not intended as medical or psychological advice, diagnosis, or a substitute for individual therapy or other professional mental health care. Reading this material does not establish a therapeutic relationship with the author.
AI tools are not therapists or mental health professionals and should not be relied upon for diagnosis, treatment, crisis support, or decisions requiring professional clinical judgement.
Everyone’s circumstances and psychological needs are different. If something you read raises concerns about your wellbeing, please consider speaking with an appropriately qualified health or mental health professional. If you are experiencing a mental health crisis or are concerned about your immediate safety, please contact your local emergency or crisis support service.
Author’s note: This reflection was written by Josie Coco with AI-assisted editing for structure, clarity, and readability. The ideas, clinical judgement, and final wording are my own.
Josie Coco is an author and Gestalt psychotherapist working with adults who are exploring the long-term effects of emotional neglect, complex trauma patterns, anxiety, depression, relational difficulty, self-worth, and life transitions. Her work is grounded in Gestalt psychotherapy, attachment theory, Polyvagal Theory and neurobiology, and a deep interest in how early relational experience shapes the body, identity, and the way we come to meet ourselves and others.
If something in this reflection speaks to your own experience, you are welcome to make a time to discover whether working together feels right.
