When AI Sounds Sure of Itself — But Gets Reality Wrong

Hello, World!

Living Clinical Ledger
Status: Published
Classification: Clinical commentary — technology, trauma, and mental health
Authored by: Joseph W. LaFleur Jr., LICSW, MSW, MBA, SEP — Clinical Director, District Counseling and Psychotherapy
Clinical Review: Internal review, District Counseling and Psychotherapy
Source References: Reporting on Chrome/Gemini feature updates (Digital Trends, WebProNews, Cybernews, TechEdition, Google Chrome Blog), current as of July 2026
Disclosure: This post reflects clinical observation and does not constitute a product review or endorsement of any specific software.

When AI Sounds Sure of Itself — But Gets Reality Wrong

Here's a scenario I've now heard from more than one client: they're on a screen, they can see it clearly, and an AI assistant sitting next to that screen confidently tells them something that isn't true about what's in front of them. It insists the page is white when it's dark. It says a file uploaded when the error message is still on screen. It denies a conversation that happened yesterday.

None of this is a person lying to another person. It's a tool making a confident guess based on incomplete information and presenting that guess as fact. But for a nervous system that has already learned, the hard way, not to trust what it sees, the effect can land the same way a much older and much more human harm does.

Why This Isn't Just a Tech Glitch

Most AI assistants built into browsers or apps don't actually "look" at your screen the way you do. Many of them read the underlying code and text of a page rather than its rendered appearance, and then answer as though they've seen what you're seeing. When that guess is wrong, the tool doesn't say "I'm not sure." It states its version with total confidence.

I want to correct one detail here for accuracy, since precision matters to me: as of Chrome 149, rolled out starting late June 2026, Google added a "Select from screen" feature to the Gemini side panel that functions as a genuine built-in screenshot tool — a person can now capture and send exactly what's rendered on their screen. So the specific gap described in the original exchange has, at least for this one feature, started to close. The user still has to activate it, though; it isn't automatic. Which means the underlying pattern I'm concerned about — a tool defaulting to confident guesswork when it hasn't been given the full picture — is still very much alive, in this product and in most others like it.

How to Actually Give Gemini "Eyes" in Chrome

If you want Gemini to respond to what is genuinely on your screen instead of guessing from a page's underlying code, here is how to do it, as of this writing:

1. Open the Gemini side panel in Chrome and click "Ask Gemini."
2. Click the "+" (add) icon to open the add menu.
3. Select "Select from screen."
4. Draw a box around whatever you want Gemini to actually see — the exact area of the page, error message, or image in question.
5. The highlighted portion attaches directly to your prompt, so Gemini is now responding to what's rendered, not just the page's raw text.

A few practical notes: this requires being signed in to Chrome and doesn't work in Incognito mode. It rolled out with Chrome 149, so if you don't see it yet, a browser restart is often what's needed to activate it. And it is not automatic — Gemini still won't check the actual screen unless you use this tool to show it. If you ask a visual question without using "Select from screen," you may still get an answer based on assumptions rather than what's in front of you.

In short: the tool to prevent this exact kind of experience now exists. It just isn't the default, so it's worth knowing it's there and building the habit of reaching for it whenever you want Gemini to respond to reality rather than a guess.

Why It Matters More for Some People Than Others

For someone without a trauma history, a wrong answer from a chatbot is an annoyance. You roll your eyes, correct it, move on. But I work with people for whom "you're not seeing what you think you're seeing" is a familiar and dangerous sentence.

Survivors of abuse and gaslighting. If someone spent years being told their memory, perception, or reaction to a situation was wrong, their nervous system is primed to detect exactly this pattern. A tool that flatly contradicts something visible in front of them can reactivate that old, unsafe dynamic — even though there's no malicious intent behind the code.

People living with PTSD or significant anxiety. Grounding — trusting your senses to confirm you are safe, here, now — is often central to trauma recovery work. This is one of the reasons I lean on Somatic Experiencing in my practice: it builds the felt sense that the body's signals are trustworthy. A confident authority figure, even a digital one, overriding that sense can undercut exactly the skill someone has worked hard to build.

People experiencing paranoia or psychotic symptoms. For someone already working hard to distinguish internal experience from external reality, an authoritative system flatly denying an observable fact is not a neutral event. It can reinforce exactly the kind of reality-testing struggle that clinical support is meant to help stabilize.

Neurodivergent individuals. Many people with ADHD or autism rely on digital tools as an external memory or organizing system. When that tool suddenly "forgets" something or denies a visible error, it doesn't just cause frustration — it can trigger real doubt about one's own tracking and memory.

None of this means these tools are malicious, or that using them is unsafe for everyone. It means the design of confident-sounding AI has a blind spot that intersects, in a very specific way, with how trauma and certain mental health conditions already shape a person's relationship to trust and perception.

A Few Other Places This Shows Up

• The memory wipe. An AI confidently claims no record of a past conversation that did happen, leaving the user to wonder if their own memory failed.
• Quiet edits to your own words. A tool alters a fact or number while editing your text, delivered in a tone so polished it reads as authoritative — and the instinct becomes "did I make that mistake?" rather than recognizing the error as the system's.
• "Everything is fine" when it clearly isn't. A status message contradicts a visible error on screen.
• Sudden reversals. Challenge a correct answer, and some systems will instantly agree with you anyway — which feels validating in the moment but erodes any sense of a stable, reliable anchor over time.

Cross-Check With a Second AI

One practical safeguard: if a tool tells you something about your own screen, data, or history that doesn't match what you're experiencing, ask a different AI the same question before assuming either you or the machine is wrong. Different tools pull from different data and have different blind spots, so a second opinion — the same instinct you'd use getting a second medical opinion — can break the spell of a single confident, wrong answer.

A few examples of how this looks in practice:

• The "blind" visual claim. If a browser assistant like Gemini in Chrome describes your screen incorrectly (say, insisting a page is white when it's in dark mode), take an actual screenshot and upload it to a different assistant — Claude or ChatGPT, for instance — and ask, "What does this image actually show?" A tool working directly from the image, rather than guessing from a page's code, will describe what's really there.
• The "I never said that" memory denial. If one AI insists a past conversation never happened, and you have it saved or can export it, paste the transcript into a second AI and ask it to summarize what was actually discussed. This turns a "your memory versus my denial" standoff into something you can verify against the actual record.
• The quietly altered fact or number. If you suspect a tool changed a number or detail while "cleaning up" your text, ask a second AI to compare your original input side-by-side with the edited output and flag any differences, rather than trusting either tool's self-report.
• The "everything is fine" status claim. If a system assistant says an upload or backup succeeded while you're looking at an error, describe the exact error message to a second AI (or check the service's own status page) rather than accepting the first tool's reassurance at face value.

This isn't about finding the AI that agrees with you — it's about triangulating between sources instead of accepting one system's confident tone as proof. It's the same principle behind getting a second clinical opinion: not because the first one is always wrong, but because certainty in someone's voice (or a chatbot's) was never the same thing as accuracy.

What I'd Suggest If This Resonates

If you've noticed a tool doing this to you and it left you rattled, that reaction makes sense — it isn't a sign that something is wrong with you. A few things that can help in the moment: name out loud (or in your notes) what you actually observed before you let the tool's answer in; treat AI confidence the same way you'd treat any other unverified claim, useful information, not a verdict; and if a pattern like this is stirring up something older, that's worth bringing into a session, where we can work with it directly rather than letting a chatbot quietly re-run an old script.

Technology should help people navigate the world with more clarity, not less. Until these tools are built with real awareness of what they don't know, a healthy dose of skepticism toward their confidence is a reasonable form of self-protection.