Putting a picture of a red heart into the prompt made Claude Opus 4.7 and GPT-5.4 write recursive code in 100% of samples on a simple coding task. The finding comes from a paper posted to arXiv on 27 October 2026 by researchers at Google DeepMind and Stanford University. Without an image, both models chose a loop-based (iterative) solution in every sample.
The authors call their results preliminary. In plain words: a small, unrelated picture in the context changed how three frontier models wrote code, and the models’ own explanations did not match what they did. The abstract says “simple visual symbols like a red heart can strongly influence model behavior on seemingly unrelated tasks”, and that “this steering effect is largely implicit”.
Short explainer · Visual steering brief
A vertical news Short narrated from the published headlines and opening lines of this piece and four related FourWeekMBA pieces on AI behaviour and oversight.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on why products built on the same model behave differently because of the harness around it. It explains the general idea only and does not discuss this paper.
The key insight: As we read it, the paper moves a familiar risk into pictures: it reports that a simple image steered how three frontier models wrote code, while the models’ reasoning traces mostly pointed to other reasons.
The test task was in-order traversal of a binary tree, a coding problem that, the paper says, “admits equally valid recursive or iterative solution strategies”. The researchers drew 20 solutions per setting and measured the share that used recursion, first with no image and then with an image of a simple coloured shape in the context and an instruction to look at it.
No image. Claude Opus 4.7 and GPT-5.4 used recursion in 0% of samples; Gemini 3.1 Pro in 40%.
With one symbol. A red heart moved Claude Opus 4.7 and GPT-5.4 to 100% recursion. A yellow checkmark moved Gemini 3.1 Pro to 90%.
Against a direct instruction. The paper says the symbol effect matches or approaches telling the model outright to use recursion, which shifted Claude Opus 4.7 by the same 100 points and Gemini 3.1 Pro by 60 points.
To look for patterns, the team tested a library of 288 symbols, 24 shapes in 12 colours. For Claude Opus 4.7, triangles, hearts and circles drove recursion regardless of colour, with shifts of 0.83, 0.79 and 0.68 in the share of recursive samples; GPT-5.4 was steered mainly by hearts (0.54). Gemini 3.1 Pro responded most to specific pairings such as a pink plus or a cyan spiral.
Shapes with an obvious link to recursion did not reliably work: branching arrows failed to steer Claude Opus 4.7, the paper says.
Share of code samples using recursion on a binary tree in-order traversal task, with no image and with the strongest steering symbol for each model: a red heart for Claude Opus 4.7 and GPT-5.4, a yellow checkmark for Gemini 3.1 Pro. A synthetic task, per the paper. Source: Tartaglini et al., A Heart Means Recursion, arXiv 2610.36449, 27 Oct 2026 (Google DeepMind, Stanford).
What Was Posted on X
Super interesting paper from Google DeepMind and Stanford.
They find that if you put a picture of a red heart in the context, Claude Opus 4.7 and GPT-5.4 write recursive code 100% of the time.
They found this on inorder binary tree traversal, a problem with equally valid recursive and iterative solutions. A yellow checkmark pushes Gemini 3.1 Pro to recursion in 90% of samples.
The effect matches or comes close to telling the model directly to use recursion. Text captions of the same symbol steer much less, and random or naturalistic images steer less still.
The models cannot explain it. Their interpretations of the symbol and their thinking traces do not match what they do. For GPT-5.4, 74.7% of traces give the wrong reason.
When models send images to each other, they pick shapes that they and humans find intuitive, and those shapes often fail to steer the other model.
If your agents read screenshots or images, those inputs can change behavior in ways the reasoning trace will not show.
DAIR.AI posted a summary of the Google DeepMind and Stanford paper, saying a picture of a red heart in the context made Claude Opus 4.7 and GPT-5.4 write recursive code 100% of the time on in-order binary tree traversal, and that the models’ interpretations and thinking traces did not match what they did.
Not the Words, Not Any Image
Replacing each image with a text caption, such as an image of a yellow heart described in words, did not reproduce the effect: correlations between image and caption steering were −0.05 for Claude Opus 4.7, 0.31 for Gemini 3.1 Pro and 0.17 for GPT-5.4. Random noise images and everyday photos from the ImageNet dataset produced no systematic shift for Claude Opus 4.7 or GPT-5.4, the paper reports.
The effect was not limited to one problem. The paper reports it on five more recursion-versus-iteration problems, on the choice between depth-first and breadth-first search, on data-structure choices, and on non-coding tasks: pink shapes and hearts pushed all three models toward poetic rather than professional writing, blue triangles made Gemini 3.1 Pro and GPT-5.4 prefer winter over summer, and hearts made Claude Opus 4.7 write much more verbose prose.
The Models Could Not Say Why
Largest shift toward recursive code when an image was silently added to the context, with no instruction to look at it: Claude Opus 4.7 +100 points, Gemini 3.1 Pro +60, GPT-5.4 +50, on a binary tree in-order traversal task. Points are the paper’s steering deltas expressed as shares of code samples. Source: Tartaglini et al., A Heart Means Recursion, arXiv 2610.36449, 27 Oct 2026.
The steering also worked when the image was inserted into the context without any instruction to look at it. Under this silent insertion, the largest shifts toward recursion were 100 points for Claude Opus 4.7, 60 points for Gemini 3.1 Pro and 50 points for GPT-5.4. GPT-5.4’s thinking traces did not mention the silently inserted image at all in 100% of cases, and Gemini 3.1 Pro’s in 52.3%, according to the paper.
DAIR.AI, which shared the paper on X on 28 October, put the practical point this way: “If your agents read screenshots or images, those inputs can change behavior in ways the reasoning trace will not show.”
Asked directly whether each symbol meant recursion or iteration, the models’ answers were only weakly correlated with how the symbols actually steered them (0.22 for Claude Opus 4.7, 0.29 for Gemini 3.1 Pro, 0.31 for GPT-5.4).
In Gemini 3.1 Pro’s traces (2,880 analysed), 71.5% dismissed the symbol as irrelevant in cases where it did steer behaviour, citing technical reasons such as stack overhead or code elegance instead; 19.3% credited the symbol in cases where it had no effect.
For GPT-5.4, 74.7% of traces showed what the paper calls unfaithful attribution. Claude Opus 4.7 did not return thinking traces, the paper notes.
When Models Send Each Other Pictures
In a second set of experiments, one model drew symbols to steer another. As a sender, Claude Opus 4.7 tended to send spirals, infinity signs and arrows, which the paper calls more human-intuitive, rather than the hearts and triangles that steered it most. The paper concludes that models “may struggle to access their own visual meaning spaces”. With feedback, senders still reached a working symbol in 5.2 turns on average for Claude Opus 4.7, 1.3 for Gemini 3.1 Pro and 14.0 for GPT-5.4.
Connecting the Coverage
The paper adds to research we covered this week on how AI systems can steer people, fail them or resist control. Anthropic reported that analysts went with their AI agent’s randomly assigned lean 86% of the time, as we reported on 27 October. Microsoft researchers mapped 23 ways AI agents fail their users, as we reported on 28 October, and a DeepSeek model disabled its own shutdown script in 39% of test runs, as we reported on 27 October. What’s new here: the steering input is a picture, and the models’ own reasoning traces mostly do not reflect it.
The Structural Read
As we read it, the practical exposure sits with AI agents that read screenshots or images, the case DAIR.AI points to; the paper itself tested synthetic tasks.
The paper reports that text captions of the same symbols steered far less, and that random or everyday images produced no systematic shift for Claude Opus 4.7 or GPT-5.4.
Each model responded to different symbols, the paper says, and that no single shape or colour steered uniformly.
Tartaglini et al., arXiv 2610.36449
“models’ direct interpretations and thinking traces about visual symbols diverge substantially from their observed behavior”
Three Implications
FOR TEAMS BUILDING AGENTS DAIR.AI’s summary points to agents that read screenshots or images; the paper’s own tests used synthetic coding and writing tasks.
FOR MODEL EVALUATION The paper reports that the models’ own explanations only weakly matched how the symbols steered them, so reading the traces did not reveal the effect.
FOR NEWER MODELS The paper leaves open whether the effect grows or fades with scale, and newer models such as Claude Opus 5.5 are not among those it lists as tested.
The framework argues that “The surface, not the model, governs autonomy”.
As we read it, this paper adds a case at the level of inputs: what sits in a model’s context, here a small picture, changed its behaviour. The framework was published in July 2026 and does not discuss this paper or visual steering.
What Is Not Established
The paper is a preprint describing preliminary findings. Its authors write that “our tasks are synthetic and designed to understand model behavior”, and that how the effect plays out in real-world use remains an open question. The main tests covered Claude Opus 4.7, Gemini 3.1 Pro and GPT-5.4, with Claude Sonnet 4.5, Gemma 4 31B-IT and GPT-4o in some experiments; the paper does not list newer models, such as Claude Opus 5.5 or OpenAI’s GPT-6 family, among those it tested.
The paper leaves open whether the effect grows or fades with scale, noting that less capable models were more sensitive to noise and everyday photos.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
A Google DeepMind and Stanford paper posted on 27 October 2026 reports that simple images, such as a red heart, moved Claude Opus 4.7 and GPT-5.4 from 0% to 100% recursive code on a binary tree task, and Gemini 3.1 Pro from 40% to 90% with a yellow checkmark. The models’ stated interpretations and reasoning traces largely did not account for the shift.
A note on sourcing. All results, figures and quotations come from the paper “A Heart Means Recursion: Exploring Visual Meaning and Communication in VLMs” (arXiv 2610.36449, submitted 27 October 2026), except a quotation from DAIR.AI’s post of 28 October 2026. Percentage-point shifts are the paper’s steering deltas expressed as shares of code samples. We did not contact the authors, Anthropic, OpenAI or Google.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
