Can You Use AI on a Game-Based Assessment?

If you have been sent a balloon-pumping game, a card-flipping task, or a screen that asks you to press the spacebar as fast as you can, you are taking a game-based assessment — and a live AI assistant will do close to nothing for you. These assessments score how you behave across hundreds of short trials, not what you write. There is no question text to paste into a model, and the thing being measured is partly your reaction time, which an assistant adds to rather than reduces.

That is the short answer. The longer answer matters because "AI-resistant" is a marketing phrase that vendors apply to very different formats, and the evidence behind it is thinner and older than the phrase suggests. This article covers what these assessments actually capture, what the vendors' own tests found, where their evidence stops, and what to do with the time you would otherwise spend looking for a shortcut.

What a game-based assessment measures

Game-based assessments come mainly from three vendors in early-career and graduate hiring: pymetrics (acquired by Harver), Arctic Shores, and the game-style reasoning tests sold by firms such as Test Partnership. Employers including BCG, Goldman Sachs and Accenture have used pymetrics in campus pipelines.

Harver's product page describes "12+ interactive, gamified experiences" taking about 25 minutes on average, measuring "cognitive, emotional, and social traits" including effort, risk tolerance, decision making, attention, focus, learning, fairness, generosity and emotion. Harver states that each assessment generates "thousands of behavioral data points" and describes them as "non-verbal behavioral-based attributes."

Those are vendor descriptions, not independent findings. But the architecture they describe is the part that matters to you: the unit of measurement is a trial, not an answer. In the balloon task, the score is the pattern of how many pumps you take before banking across dozens of balloons. In a go/no-go task, it is the distribution of your response latencies and your error type — whether you miss targets or press on non-targets. A correct answer does not exist for most of these. Harver's own framing, in a post published 13 October 2026, is that "assessments that measure behavior in the moment are harder to fake, because there is no correct answer to rehearse."

Notably, the same post declines to overclaim: "No responsible assessment scientist should claim that any hiring method is completely 'uncheatable.'" That is an unusually honest line from a vendor, and it is the right frame for everything below.

Why an AI assistant does not transfer to this format

An interview assistant helps when the bottleneck is knowledge or articulation — recalling a system design pattern, structuring a behavioral answer, reading an unfamiliar error message. On a game-based assessment, none of those is the bottleneck.

Three things break the usual workflow:

There is nothing to copy. The tasks are graphical and dynamic. The only text on screen is the instruction panel, which explains the rules but contains no problem to solve.

Latency is scored, not just tolerated. Reading an overlay, deciding, and returning to the task costs several hundred milliseconds per trial. On a task where response time is itself a measured variable, that cost lands directly in the score.

The scoring is a profile, not a total. Models are built by comparing your trial-level pattern against a reference group. Shifting your behavior mid-assessment tends to produce an internally inconsistent profile rather than a better one.

This is the same reasoning that applies to any format where the measured quantity is behavior rather than output. It is a different problem from the one in what AI-assisted coding interviews actually grade, where the artifact you produce is the thing under review.

What the vendors actually tested, and when

Arctic Shores is the only vendor in this space that has published an attempt to break its own product, which makes its results worth reading carefully — including their limits.

The company tested its task-based assessment against generative AI with help from two UCL postgraduate researchers, trying three approaches: scanning the instruction screens to text and pasting them into ChatGPT; uploading task images to Google's Bard, because ChatGPT could not read images at the time; and commissioning an engineer to build an OCR-driven bot. None completed the assessment. ChatGPT rephrased the instructions without adding useful guidance; Bard struggled with multi-element tasks and in one case hallucinated a star that was not among the available shapes.

That test is dated. Its page metadata records a last update of October 2023. The central constraint it describes — that the model could not see the screen — no longer holds. Every frontier model released since then reads images natively, and screen-aware agents are widely available. The 2023 result is real evidence that text-extraction attacks failed against a visual task; it is not current evidence about what a 2026 multimodal model can do with the same screen.

Arctic Shores' candidate help centre takes a narrower line and does not rely on that test. It says employer policy governs, that "it's a good idea to check directly with the organisation you're applying to," and that in the company's testing "none of them could complete an Arctic Shores assessment successfully" — with no claim about detection methods or penalties.

The contrast with the same company's other experiment is the most useful thing in the published record. Testing a retired text-based reasoning assessment against a sample of 36,000 early-career candidates, Arctic Shores found GPT-4 with chain-of-thought prompting scored 21 out of 24 on verbal reasoning, "better than 98.8% of human candidates." On numerical reasoning the same model performed far worse, with only two of five prompting styles edging above the human average.

Read those two studies together and the conclusion is about format, not about assessments as a category. Text-based verbal reasoning was close to fully solved by a 2023 model. Visual, timed, trial-based tasks were not. "AI-resistant" describes a specific design property — no extractable text, timing in the score — not a property of anything an employer calls an assessment.

Where the vendor claims outrun the evidence

Test Partnership markets AI-resistance as a design choice, saying it builds "AI-resistance into the test format itself, rather than relying on surveillance," and that its game-based product is "AI-resistant by design" because there is no static question text to copy. Its static reasoning product relies instead on showing the stimulus before a rapid run of questions, on the theory that there is too little time to paste each item into a model.

The page offers no empirical study for either claim, and no data on how often candidates actually use AI on assessments. Harver's figures — 100 million assessments matched, a 98% completion rate, validation "for fairness across gender, ethnicity, and socioeconomic status" — are likewise self-reported, with no bias audit published alongside them.

None of this means the vendors are wrong. The structural argument is sound and the Arctic Shores test was a real attempt at falsification. It means the honest state of the evidence is: plausible mechanism, one dated experiment, no independent replication.

The policy question is separate from the capability question

Whether a tool would help and whether you are permitted to use it are different questions, and the second one is the one with consequences.

Most game-based assessments arrive as part of a graduate or early-career pipeline, where AI-use rules are usually written into the application terms rather than shown on the assessment screen. Those terms are where misconduct findings come from, and a finding can follow you across a firm's future cycles — the mechanics are covered in what counts as AI misconduct in a graduate scheme.

Arctic Shores' own guidance points the same way: check with the organisation, because the employer sets the rule.

What to do instead

Preparation genuinely has limited value here, and the vendors say so — Arctic Shores states the assessment "needs no preparation or practice" because it measures natural responses. That is partly self-serving, but it is close to true for the behavioral tasks. The practical steps are mundane:

• Do the practice rounds the platform offers. Familiarity with the interface removes the first-trial confusion that otherwise shows up as noise in your profile.
• Use a wired mouse and a machine that is not thermally throttling. Input lag is indistinguishable from slow reaction time.
• Do it once, alert, without interruption. Partial completion and long mid-task gaps are visible in trial-level data.
• Do not try to present a persona. The scoring compares your pattern to a role model; a deliberately "bold" or "cautious" performance usually reads as inconsistent rather than ideal.

If you are flagged or your result is questioned, the process from there — what gets reviewed, what you can ask for — is the same one described in what happens when a coding assessment flags you.

Where this leaves a desktop assistant

Control is a desktop AI interview assistant for Windows and macOS, built for interviews and assessments where the work is spoken or written: live transcription of the interviewer, screenshots of a problem, and answers in an overlay. That workflow fits a system design discussion or a coding screen. It does not fit a balloon task, and we would rather say so than imply otherwise.

If what you are facing is a game-based assessment, the useful answer is that there is no shortcut worth taking, and the vendor evidence — dated as it is — points the same way. If what you are facing next in the same pipeline is a technical screen or a live interview, that is a different format with a different answer, and our guide to AI coding assessment workflows is the place to start.