ChatGPT and other AI chatbots can often identify an item from a clear photo and explain what affects its value, but their prices are not reliable on their own, as of October 2026. A chatbot can state a confident figure with no sale behind it, or repeat prices from old articles. Use it to name the item and list what to check, then confirm the value with sold listings.
That is not a knock on chatbots. They are general tools, and valuing a used item depends on two things they are weakest at: seeing small details that decide the exact version, and knowing what that version sold for last month. With the right prompt and a quick check of sold prices, they become genuinely useful.
Quick answer
- Chatbots are good at: naming the item, reading visible marks, explaining what makes some versions valuable, and suggesting search terms.
- They are weak at: current prices, condition grades, small marks in poor photos, authenticity and rare variants.
- Common failure modes: invented prices, stale prices, the price of the best example applied to yours, and confident answers from blurry photos.
- Better prompts help: ask it to identify first, state its confidence, list what would change the value and give sold-listing search terms.
- For anything that could be worth thousands, a specialist or a human expert is the better choice.
What does ChatGPT do with a photo of an item?
It describes what it sees and matches that to what it learned from text and images, then answers in fluent prose. For an item with a clear name and a visible mark, such as a Pyrex bowl with its pattern, a coin with a readable date or a watch with its brand on the dial, the identification is often right. Ask "what is this worth" and it will usually give a range and a list of factors.
Unless it runs a web search, which current versions often do on their own for price questions, it does not look up recent sales of that exact item. Without search, its price comes from patterns in its training data, which can include old price guides, collector articles that quote record sales, and listings that never sold. Even with web search on, it may cite asking prices or articles rather than sold results.
Why do chatbots get prices wrong?
The main reason is that language models are built to give an answer, and a price sounds like an answer. Researchers at OpenAI and Georgia Tech put it plainly in a September 2025 paper, "Why Language Models Hallucinate": "Like students facing hard exam questions, large language models sometimes guess when uncertain," producing plausible but incorrect statements instead of admitting doubt. The paper argues that common evaluations reward guessing over saying "I don't know." ChatGPT's own interface carries a standing note that it can make mistakes and that important information should be checked.
The common failure modes when valuing items:
| Failure mode | What it looks like | How to catch it |
|---|---|---|
| Invented price | A specific figure with no source | Ask for the sold listings behind it; check eBay Sold Items yourself |
| Stale price | A figure from a peak market years ago, such as 1990s collectibles | Check sales from the last few months |
| Best-example price | A record or mint-condition figure applied to a used item | Ask for the price of your condition specifically |
| Wrong version | Right object, wrong year, mark, edition or size | Read the mark yourself and compare |
| Overreading a blurry photo | A confident date or mark it cannot actually see | Ask it to say which details it could not read |
| Silver or gold at an old spot price | Melt values from a different year | Recalculate with today's spot price |
How do you prompt ChatGPT to value something better?
Ask it to identify first and price second, to say what it cannot see, and to give you the search that settles it. A prompt like this works far better than "what is this worth":
- Give it good photos. The front, the back or base, and every mark close up. One item per conversation.
- Tell it what you know. Size, weight, material, where it came from, any writing you can read better than the camera.
- Ask for the identification and its confidence. "Identify this item as exactly as you can: maker, model or pattern, approximate date. Say how confident you are and which details you could not read."
- Ask what would change the value. "What versions or details make this kind of item worth more or less, and which of those can you see in my photos?"
- Ask for search terms, not just a number. "Give me the exact eBay search to find sold examples that match."
- If it can search the web, ask for sold results with dates. "Find recent sold prices, not asking prices, and list the sources and dates."
- Then check yourself. Run the search, filter to Sold Items, and take the middle of five or more true matches. Our what is it worth guide explains this step.
A worked example
Say you photograph a porcelain figurine of a girl with two geese from the front and ask a chatbot what it is worth.
It may correctly identify it as a Goebel Hummel "Goose Girl" and give a wide range, perhaps a few hundred dollars, drawing on older articles from when Hummels peaked. The trouble is that the front of the figurine does not carry the information that sets the price. The trademark on the base does.
So turn it over and add a photo of the base. Suppose it shows "Goebel" with no bee: that is the Missing Bee mark, used from about 1979 to 1990. As our Hummel figurines value guide shows, figurines with that mark typically sell for about $10-50, while the same design with the 1950s Full Bee mark brings about $50-250. Ask the chatbot again with the base photo, ask for the sold-listing search, and check eBay yourself. The honest answer for a Missing Bee Goose Girl in good condition is tens of dollars, not hundreds. The chatbot helped name it. The base photo and sold listings priced it.
When is a specialist tool or a human better than a chatbot?
A specialist tool is better when the category has its own marks and markets, and a human is better when money or authenticity is at stake.
| Situation | Better choice | Why |
|---|---|---|
| Quick "what is this" for an everyday item | Chatbot or Google Lens | Fast, broad identification |
| Coins, cards, watches, bags | Category guides, specialist apps, sold listings | The exact version and condition decide price |
| Item could be worth thousands | Dealer, auction house specialist or grading service | They examine the item and know current buyers |
| Authenticity matters (designer goods, signatures, rare coins) | Brand, authentication service or grading service | Photos cannot prove a careful fake is fake |
| Insurance, estate or tax donation | Qualified appraiser | Needs a signed document; IRS rules apply over $5,000 |
On coins specifically, even purpose-built apps struggle with grade. A November 2025 review in the American Numismatic Association's Reading Room found CoinSnap's identification "excellent" but its grading "unreliable," and gave an example of one proof half dollar valued at $7.50 to $75. A general chatbot has the same limits and less category data. For coins, start with our coin value guide and the coin grading scale.
Common mistakes
- Taking the first number. A chatbot's first price is a starting point, not a value.
- Sending one photo of the front. The base, back or inside usually carries the mark that sets the price.
- Not asking what it could not see. Chatbots rarely volunteer uncertainty; ask for it.
- Accepting a range built on asking prices. Ask specifically for sold results with dates.
- Using it to authenticate. A chatbot saying an item "appears genuine" is not authentication.
- Assuming web search fixes everything. Search helps, but the results still need to be the exact version and sold, not listed.
How WorthPic differs from a general chatbot
WorthPic runs on the newest and most capable AI model available, tuned for exactly this kind of identification. Unlike a general chatbot, it is built for one question: what is this item and what is it worth. You pick one of twelve categories and upload one to four photos, and it reads the marks, stamps, dates, labels, model numbers, materials and visible condition, identifies the item, and gives an honest US dollar range with a most likely figure and what drives the price. When the price depends on something the photos do not show, it says which extra photo would settle it, such as the base of a figurine or the reverse of a coin. It cannot know weight, metal purity, hidden damage or whether a careful fake is genuine. It identifies and estimates. It does not authenticate, grade or appraise.
What to photograph before asking any AI
- The whole item in daylight, square on, no flash, plain background.
- The base, back or inside, where marks usually are.
- Each mark or signature close up and in focus.
- Damage or repairs, so the condition is clear.
- A ruler or coin for scale, since size often separates versions.
When to get it looked at in person
If a chatbot or your own sold-listing search suggests an item could be worth thousands, or the value depends on authenticity, metal, stones or grade, take it to a person. A coin dealer or PCGS and NGC for coins, a jeweler for gold and stones, an auction house specialist for art and antiques, a grading company for cards, and a qualified appraiser when you need a document. Our guides to AI appraisal and online appraisal explain the steps between a chatbot and a formal appraisal.
Sources
- Kalai, Nachum, Vempala and Zhang, Why Language Models Hallucinate (arXiv, September 2025): why models guess when uncertain.
- ANA Reading Room, Mind the App: Coinsnap's Pros and Cons: independent test of photo-based coin identification, grading and pricing.
- eBay Product Research announcement: three years of eBay sales data available to sellers.
- IRS Instructions for Form 8283: qualified appraisal requirement above $5,000.