Why AI Misreads Price Levels on Charts

AI misreads chart price levels because it never sees a price. It estimates each level by reading small axis text and interpolating vertically between gridlines, so anything that distorts that chain, image compression, a logarithmic axis, a cropped price scale, an overlay covering a wick, or a rounded label, moves the number it hands back. Here are the seven causes and the fix for each.

How does an AI actually arrive at a price level?

Understanding the mechanism makes every failure below obvious. The model receives an image, not an OHLC series. To say "resistance at 187.40" it has to do three things in sequence: find the axis labels and read them as text, work out the pixel-to-price mapping between two of them, then locate the relevant candle wick in pixel space and convert its vertical position back into a price.

Every step in that chain is lossy. Axis labels are small text, which OpenAI's vision documentation flags as a known weakness. The mapping assumes linear spacing. The wick location assumes the wick is visible and not overlapped. Errors compound, and the model reports the result with the same fluency it reports the trend, which is why a wrong level does not look wrong.

1. The price axis is cropped or absent

The most common cause and the easiest to fix. Traders crop screenshots to make them look tidy, and the right-hand price column is the first thing to go. Without it the model has no calibration at all, so it either refuses, or, more often, infers plausible prices from the ticker it recognised and the range it assumes that instrument trades in. That inference can be off by an order of magnitude.

The fix. Always include the price axis, uncropped, with at least three labels visible. If your platform hides labels at narrow widths, widen the window before capturing.

2. Resolution and compression destroy the numbers first

Axis text and one-pixel wicks are the highest-frequency detail in the image, and high-frequency detail is exactly what downscaling and lossy compression discard. A phone photo of a monitor adds moire, glare and keystone distortion on top. The model does not report that it could not read the label; it reads the closest thing to a label it can find.

The fix. Native screenshot, full resolution, PNG where the tool accepts it. Do not resize down to meet an upload limit if you can crop horizontally instead. If the axis text is hard for you to read at 100 percent zoom, it is worse for the model.

3. Logarithmic scales break the interpolation

On a linear axis, equal vertical distance is equal price change, and interpolating between two labels is arithmetic that works. On a logarithmic axis, equal vertical distance is equal percentage change, so the same interpolation is wrong, and the error grows with the vertical gap between labelled gridlines. This bites hardest on long-horizon crypto and equity charts, which are exactly the charts most likely to be set to log.

The fix. Switch to linear for the screenshot when the visible range is not huge. If you need log, say so in the prompt and expect wider tolerance on any level in the sparse regions of the axis.

4. Overlapping indicators cover the evidence

A level is defined by where price reacted. If a moving average ribbon, a Bollinger band, a VWAP line and two drawn trendlines run through the same region, the wick the model needs is behind three other lines of similar colour. Documentation specifically calls out charts where solid, dashed and dotted styles carry different meanings as a weak case, and an indicator-heavy chart is that case at maximum density.

The fix. Remove every overlay you are not asking about. If you want a read on structure, send a clean candlestick chart. If you want a read on an indicator, send that indicator and nothing else competing with it.

5. Cropped and clipped candles

Two versions of the same problem. The first is a horizontal crop that removes the swings which created the level, so the model marks a level from the reactions that remain and misses the one that matters. The second is a vertical clip, where a spike runs off the top or bottom of the visible range and the model reads the truncated wick as the extreme. Both produce a level that is internally consistent with the image and wrong about the market.

The fix. Include roughly one hundred to two hundred bars, and check that the visible price range contains the full high and low of the structure you are asking about. Zoom out before you capture, not after you get an answer you dislike.

6. Rounded, abbreviated and thinned axis labels

Charting platforms compress axis labels to fit: 1.2K, 48.5k, 1,850 with the decimals dropped, or every other gridline unlabelled at small window sizes. The model reads what is printed. If the axis says 1.2K where the true value is 1,247, the interpolation inherits that rounding, and a level quoted to two decimals from a thousand-rounded axis carries false precision.

The fix. Widen the chart pane so the platform prints more granular labels, and be suspicious of any returned level with more decimal places than the axis actually shows. That mismatch is a reliable tell that the number was interpolated rather than read.

7. Missing timeframe and missing ticker

Not a numeric error but a framing one, and it changes which levels come back. The same instrument has different significant levels on a five-minute and a daily chart. If the timeframe label is not in the image, the model infers it from candle density and volatility, and when it guesses wrong it applies the wrong sense of what counts as a meaningful reaction. A missing ticker has a subtler effect: the model may anchor on price ranges typical of an instrument it thinks it recognises.

The fix. Keep the platform's ticker and timeframe header in the capture, and state both in text anyway. Redundancy is free here.

Why do the levels change between two runs of the same chart?

Because support and resistance is a judgement, not a measurement, and the model re-makes that judgement from scratch every time. There is no canonical definition of how many touches make a level, how wide a zone is allowed to be, or whether a wick counts, so two reads can both be defensible and different. Add the ordinary variability of language-model output and you get level drift on identical input.

Consistency is therefore a property worth testing rather than assuming. Upload the same image twice, some minutes apart, and see whether the structural read holds and how far the numbers move. A direction that flips on identical input is a real finding about the tool. Our public evaluation protocol makes repeat-submission consistency a scored dimension for exactly this reason; the page publishes the rubric and the test-set rules, not scores for named products.

Where Bullynx sits on this, honestly

Bullynx reads screenshots, so every mechanism above applies to us as well. Our level numbers are visual estimates derived from your image, and they can be wrong when the axis is cropped, the capture is compressed, the scale is logarithmic, or the wick is buried under an overlay. We do three things about it rather than claiming it away. The output always names the levels as zones tied to a visible reaction, so you can check the evidence rather than the number. The read always includes an invalidation level, which is the field where being approximately right is least dangerous, because you size around it. And the routine is the same on every upload, so when a level moves between two reads of the same chart, that is information rather than noise. What we cannot do is see outside your crop, read an illegible axis, or know that the level was formed on news. Those stay yours.

The practical rule

Use an AI chart read for the labelling, not for the numbers. Structure, sequence, which side the trend is on, whether a level has been respected more than once: those are shape questions and the answers are usually sound. The exact prices are the weakest field in the output, and they are also the field people copy straight into a broker. Confirm every level on your own chart before it becomes a stop or an entry, and treat overconfident numeric claims about accuracy from any vendor with the scepticism regulators recommend for AI investing claims generally.

For the wider set of blind spots beyond price precision, see AI chart analysis: 7 things it gets wrong. For what a good screenshot looks like in the first place, see can AI read a chart screenshot. And if you want the workflow end to end, the chart screenshot analyzer page shows what a full read contains.