I'm becoming AI-blind: I'll read a clean paragraph or look at an ordinary photo of a stranger and genuinely can't tell if it's real anymore. That's not a personal failing — it's what happens when generative models get good enough to erase the tells your eye used to catch without even trying.
Short answer: "AI-blindness" is losing the ability to reliably tell AI-made content from real content — or the reverse, doubting real content on a hunch. Fix it by testing your instincts against known answers instead of trusting your gut, checking an image's provenance metadata when it exists, and using a detector as a second opinion, never a final verdict.
I test AI writing, image, and detection tools most weeks for this site, and the pattern I keep running into isn't that I'm bad at spotting AI content — it's that my confidence and my accuracy have come apart. A September 2025 study by Conjointly found something similar at scale: 301 US adults shown real photos next to images made with Google's Gemini 2.5 Flash scored close to chance (49% on real photos, 52% on AI ones), yet 42% said they felt confident or very confident telling the two apart — up from 31% in June 2023. The gap between how sure people feel and how accurate they actually are is the real problem, and it's the one this guide is built to close.
What you'll need
Nothing to install to start. You need ten minutes, a mix of images or text samples you already know the true source of (your own camera roll works for photos), and a willingness to score yourself honestly instead of assuming you're fine. If you want tool-assisted checks later in this guide, having a free account with one AI detector helps, but it's optional for the first two steps — the self-test matters more than any tool.
Step-by-step: fixing AI-blindness
1. Name which failure mode you actually have
"AI-blind" covers two different problems, and they need different fixes. One is missing real AI content because it's too convincing — you get fooled. The other is flagging real content as fake because you've grown suspicious of everything — you distrust people who did nothing wrong. In my testing, most people who say they're "becoming AI-blind" actually have some of both, which is why a single tip like "look for extra fingers" stopped working months ago.
2. Run a blind self-test with known answers
Pull five photos you took yourself (so you know they're 100% real) and generate five with any image tool — Midjourney, Gemini, whatever you have access to. Mix them, look at each for five seconds, and guess before you second-guess. Score yourself. When I ran this on myself with a fresh batch of Gemini images, I got 6 of 10 right — barely better than a coin flip, and worse than I expected going in. That number, not a vague feeling, is your actual starting baseline.
3. Check provenance metadata before you eyeball anything
Many images now carry Content Credentials — a tamper-evident record of what made them and what's been edited since, built on the C2PA standard that Adobe, Google, Microsoft, and others back. Before you play detective on pixels, drop the file into Content Credentials Verify, a free tool with no account required. If credentials are attached and unbroken, you have a straight answer; if there's nothing attached, that tells you nothing either way, since plenty of real cameras and older AI tools never add it.
4. Use a text detector as a tiebreaker, not a verdict
For written content, my full process — the specific rhythm and phrasing tells, plus where detectors get it wrong — is in how to spot AI writing. The short version for this guide: read for yourself first, run a detector like Originality.ai second, and never let one score override your own read of the context.
5. Watch for the video and audio tells that still hold up
Deepfake video and cloned voice are the hardest category, but a few tells still work more often than not: lighting on a face that doesn't match the room behind it, blinking that's slightly too regular or absent, and audio that's a hair out of sync with mouth movement in a call or clip. None of these are proof on their own — treat them as reasons to slow down, not reasons to accuse.
6. Re-test yourself every few months
The models generating this content change every few months, and a self-test from six months ago tells you nothing about your accuracy today. Repeat step 2 with fresh AI-generated samples on a recurring basis — I do it roughly every quarter — and expect your score to drop again each time the underlying models improve, because that's exactly what happened to the Conjointly respondents between 2023 and 2025.
Example prompts you can copy
Use these when you want a second read, not a replacement for your own judgment:
- Ask a model to argue both sides: "Give me three specific reasons this could be real and three specific reasons it could be AI-generated: [describe the image or paste the text]."
- Stress-test written content: "Rate this paragraph 1–10 on how likely it is to be AI-written, and name the specific sentences driving your score: [paste]."
- Verify a claim independently: "I can't tell if this photo or quote is real. What outside sources — reverse image search, original publication, primary source — could confirm it?"
- Check your own drift: "I've attached my last 10 guesses on real-vs-AI images with the correct answers. What pattern in my wrong guesses stands out?"
Common mistakes to avoid
The biggest one, backed by the Conjointly numbers above, is trusting confidence over a tested baseline — feeling sure is not the same as being accurate, and the gap between the two has been growing, not shrinking. Second is testing once and assuming that score holds; it doesn't, because the generating models keep improving. Third is treating a single detector score as a verdict instead of one data point among several. Fourth is over-correcting into suspicion of everything, which is its own kind of AI-blindness — I've caught myself doubting a friend's real vacation photo for no reason beyond "it looked a little too nice." Fifth is skipping provenance checks entirely and going straight to guessing by eye, when a free tool can sometimes just answer the question directly.
Tools that make this easier
None of these tools replace the self-test in step 2, but they're useful once you know your own baseline. For text, Originality.ai runs $14.95/month on its Pro monthly plan and is built for agencies scanning at volume; AI or Not covers text, images, and short video checks from a free tier (1 million words and 20 image checks a month) up to a $5/month Pro plan with 500 image checks and 1,000 seconds of video. Hive Moderation offers a similar detection API for platforms screening uploads at scale, but pricing is quote-only — you'll need to contact their sales team. For a deeper look at how any of these actually hold up under testing rather than vendor claims, see AI tool reviews and AI tool ratings; if you're regularly producing AI-assisted writing yourself, best AI writing tools and how AI text watermarking works cover the other side of this problem, including how Claude’s own watermark works.
| Tool | Best for | Entry price | Free tier |
|---|---|---|---|
| AI or Not | All-in-one checks: text, images, short video | $5/month (Pro) | Yes — 1M words + 20 image checks/month |
| Originality.ai | Agencies/publishers scanning text at scale | $14.95/month (Pro, monthly) | No permanent free plan |
| Hive Moderation | Platforms screening user uploads at scale | Custom (contact sales) | No |
| Content Credentials Verify | Checking an image's built-in provenance metadata | Free | Yes — no account needed |
My take
The fix for AI-blindness isn't a sharper eye — nobody's eye is sharp enough for this generation of models anymore. It's a tested baseline instead of a felt one, a provenance check before a guess, and a detector used as one input among several instead of a verdict. When I started scoring my own guesses instead of trusting my gut, the number that came back was lower than I expected, and that was more useful than another list of "5 signs it's AI" that stops working the next time a model ships.
Frequently Asked Questions
Is it free to check if I'm becoming AI-blind?
Yes. The self-test in step 2 costs nothing beyond photos you already have, and Content Credentials Verify is free with no account. Paid detector tiers mainly add higher volume and batch scanning, not better core accuracy for casual use.
How long does it take to fix AI-blindness?
The self-test takes about ten minutes and gives you an honest baseline immediately. Building the habit of checking provenance and using a detector as a second opinion takes a few real attempts — most people get comfortable after doing it two or three times.
What is the easiest way to tell if something is AI-generated?
Check for Content Credentials first if it's an image — that's a direct answer when it's present. If nothing's attached, read for specific, checkable detail rather than "vibes," and run a detector as a tiebreaker rather than a first opinion.
Can AI content detectors be wrong?
Yes, regularly, in both directions — they miss AI content and flag real content. That's exactly why the self-test and provenance checks in this guide matter more than any single tool's score.
Will I ever be as accurate as I used to be?
Probably not against the newest models, and that's the honest answer — accuracy has dropped across the board as generation quality improved, per the Conjointly data above. The realistic goal is knowing your actual baseline and layering checks, not restoring some earlier level of certainty.