The Solver Threat: AI Math Assistance in 2026 and What Actually Resists It

Findings

1. Consumer math solvers: capability and reach

Photomath combines a computer-algebra system with OCR (extended to handwriting recognition since 2016) to scan a printed or handwritten problem — including word problems — and produce a step-by-step solution in seconds, spanning “elementary through college” math: arithmetic, algebra, geometry, trigonometry, statistics, and calculus [4][5]. Scale is large: as of 2021, over 220 million downloads and roughly 2.2 billion problems solved per month [4]. Symbolab explicitly advertises acceptance of “written pages and screenshots” alongside typed notation and natural-language queries, covering pre-algebra through calculus, trigonometry, physics, and statistics, framing output as step-by-step rather than answer-only [6]. Wolfram|Alpha’s strength is symbolic/closed-form computation with free-form input — very strong on canonical algebra, calculus, and equation-solving, historically weaker on problems needing semantic parsing of an ambiguous word problem. Camera-first apps such as Gauth (and Microsoft Math Solver, which adds handwriting recognition and graphing to a similar OCR-plus-solver pipeline) follow the same pattern: camera in, stepped answer out in seconds, several with a live human-or-AI tutor chat layered on for anything the automatic solver does not parse cleanly. The common thread is that these tools are strongest exactly where school and practice-app problems are weakest by design necessity: single, well-posed, closed-form problems with one correct final answer.

2. General AI assistants: from AIME to IMO gold

Frontier LLMs have moved through and past the competition-math tier that used to be a meaningful ceiling. AIME (a US qualifying olympiad exam) is now listed by at least one benchmark tracker as an “archived” or saturated benchmark because “performance on this benchmark has saturated” and providers have stopped running new releases against it, with a top score of 98.12% recorded for Gemini 3.1 Pro Preview and a note that nine of the ten top-ranked models on the leaderboard are reasoning models [12]. Competition-grade natural-language proof generation, the harder frontier, has also fallen further than expected. In 2024, DeepMind’s AlphaProof and AlphaGeometry 2 reached 28/42 points (silver-medal standard) at the IMO, but only after problems were manually translated into the formal language Lean, and the hardest problem took up to three days of compute [3]. One year later, in July 2025, an advanced version of Gemini with “Deep Think” reached 35/42 (gold-medal standard), solving five of six problems perfectly, end-to-end in natural language — no formalization step — within the same 4.5-hour-per-session limit human competitors face [2]. The IMO’s own review confirmed the submitted solutions were “complete and correct” but stated “their review does not extend to validating our system, processes, or underlying model” [2] — these are graded artifacts, not audited systems, and the result still involved substantial human curation of training data and general problem-solving hints [2]. OpenAI separately reported comparable gold-medal-level results for an experimental model around the same period, though not run through the same official IMO coordination process; treat that claim as directionally consistent with DeepMind’s verified result rather than an independently audited score. Where general models still fail is the genuine research frontier: Epoch AI’s FrontierMath sources problems from professional mathematicians across number theory, real analysis, algebraic geometry, and category theory that take experts multiple hours each, with a top “Tier 4” tranche taking several days of expert effort — designed to keep testing frontier systems as easier benchmarks saturate [11]. That is well above anything a K-12/university practice app would assign, but it matters for Math Challenge’s ambition to reach PhD-level content: somewhere between “hard problem set” and “open research problem,” solvers stop being reliable, and that boundary is worth knowing precisely before designing top-tier content.

3. Multimodal screenshot-to-answer: why cameras defeat in-app defenses

The mechanism that makes most in-app anti-cheat design irrelevant is simple: a photograph of a screen is an out-of-band channel. It never enters the app’s DOM, never fires a keyboard or paste event, never touches its network requests, and is invisible to any JavaScript-level defense (disabling copy/paste, obfuscating source, blurring on tab-switch, watermarking, even most in-browser proctoring). The photo leaves the device through the camera and a second app or device — a surface the practice app cannot instrument at all. Every consumer solver above is built around exactly this workflow — camera in, stepped answer out, in seconds [4][5][6]. The newer wrinkle for 2026 is that this no longer needs a discrete “photo, wait for OCR” step: live multimodal assistants (screen-share or continuous camera/vision modes across GPT, Gemini, and comparable products) let a student share a screen in real time and get a spoken answer conversationally, including narration of what to do next — reading a problem off a moving or partially obscured screen, mid-interaction. No purely client-side or purely time-based defense is durable on its own; the two responses that do not depend on “can the student get a photo out” are (a) requiring a submitted artifact that is not a single reproducible string — a process, a UI manipulation state, a dialogue — and (b) making one external round-trip cost more, cumulatively, than the points it earns, via adaptive re-asking (§6 and Design Implications below).

4. How widespread AI use already is among students

This is not a marginal behavior. HEPI’s December 2024 survey of 1,041 UK full-time undergraduates (fielded by Savanta) found 92% had used any AI tool and 88% had used generative AI specifically for assessments, up sharply from 53% a year earlier; time savings (51%) and perceived quality improvement (50%) led stated motivations, with fear of cheating accusations (53%) and hallucination risk (51%) the leading deterrents — not a lack of capability or awareness [1]. Pew Research’s January 2025 release (Ipsos, Sept–Oct 2024, n=1,391 US teens 13–17) found 26% had used ChatGPT for schoolwork, double the 13% recorded in 2023 — and teens’ own judgment of acceptability varies sharply by task: 54% call it acceptable for researching a topic, but only 29% for solving math problems (28% call it unacceptable), and 18% for writing essays [10]. That asymmetry matters here: math is the subject students themselves are most divided on, a real opening for a product whose scored/ranked surfaces can credibly claim to resist casual solver use. Usage among people already given AI access skews toward direct answer-seeking: Anthropic’s analysis of roughly 575,000 academic Claude.ai conversations found Computer Science and STEM heavily overrepresented relative to enrollment (Computer Science alone 36.8% of conversations against 5.4% of US degrees), and close to half (~47%) of conversations were “Direct” — seeking an answer or finished content with minimal back-and-forth [9]. Task breakdown skewed toward higher-order work (Creating 39.8%, Analyzing 30.2%) and away from recall (Remembering 1.8%) [9], consistent with a “let the tool reason, I’ll take the output” default — precisely the behavior a practice app needs to make unrewarding.

5. The failure of AI-text detectors, and what it implies

Two lines of published research undercut the idea that a detector can gate cheating reliably. First, detectors are demonstrably biased: GPT-detector studies find they “consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified,” and the same simple prompting strategies that reduce this bias also let a user evade detection entirely — the same lever cuts both ways [7]. Second, detectors are fragile to adversarial evasion generally: research testing detection against ChatGPT and Claude text found paraphrasing, random spacing, and adversarial perturbations “can significantly diminish detection effectiveness,” concluding current methods lack robustness against even unsophisticated evasion [8]. The implication here is direct: any “explain your reasoning” field cannot safely use an AI-detector as an automated pass/fail gate — doing so would be trivially evadable and would risk disproportionately flagging genuine work from Spanish-first or otherwise non-native writers in a product whose own requirements mandate bilingual UX. Detection, where used at all, belongs as a soft signal feeding human review, never an automated block.

6. Design responses that actually survive contact with a solver

The throughline across every mitigation that holds up is the same: stop scoring a single, reproducible final artifact, and start scoring something a solver cannot hand over in one shot — a process with graded intermediate steps, a judgment about someone else’s work, a live adaptive follow-up, or a manipulated UI state. None of this makes a determined student solver-proof; it raises the round-trips, translation work, and time cost per point earned — the only lever a self-serve PWA actually controls. The table and Design Implications below turn this into a concrete build catalogue.

What survives a solver

Item format	How easily a solver defeats it	Gradability
Plain final answer (“solve for x”)	Trivial — seconds, near-total success [4][5][6]	Fully auto-gradable; weakest format
Single-step word problem	Easy for multimodal LLMs; OCR-only solvers weaker but closing	Auto-gradable with a parser
Multi-step word problem, named sub-quantities	Still defeated, but requires transcribing the whole problem	Auto-gradable per step; friction, not immunity
“Show your work” / full process	Solver generates a complete process to copy verbatim	Needs human/AI grading; copied text not reliably detectable [7][8]
“Spot the error in this solution”	Harder — solver must evaluate an argument, not just produce one	Auto-gradable (which line, what error)
Estimation / order-of-magnitude only	Weak alone — an exact solver answer trivially satisfies a range check	Easy to auto-grade; pair with justification
Interactive manipulation (drag point, build graph, balance equation)	Solver can describe the target, but performing the UI action still needs the student	Graded on final UI state, not a string
Multi-answer / select-all, misconception distractors	Moderate — brute solving gets the set, but targeted distractors weaken the shortcut	Fully auto-gradable
Adaptive follow-up (new numbers, same method)	Strong — catches “answered once” vs. “can repeat”; defeated only by re-querying every time	Auto-gradable, fully app-controlled
Socratic in-app tutor dialogue	Strong vs. static screenshots; degrades if solver text is pasted in	Needs its own grader — same arms race, one layer up
Live oral defense / synchronous check	Very strong	Needs live staff/infra; poor fit for a self-serve PWA
Original/novel research-level proof	Resists solvers and general LLMs at the true frontier [11][2][3]	Not gradable at scale; expert review only

Design implications

1. Default hint/challenge format: show a worked solution with one wrong step, ask which line and why — evaluating an argument beats producing one.
2. Replace the single-answer box with a structured multi-field process (each operation plus its intermediate result), graded per step, so a copied final answer without matching steps fails automatically.
3. Mandatory adaptive re-asking: follow a correct answer immediately with an isomorphic problem (same method, new numbers); “right once, wrong on the variant” is a real signal, especially in ranked/leaderboard modes.
4. Multi-answer/select-all items with distractors built from documented, topic-specific misconceptions, not generic “close numbers,” so brute-solving is a weaker shortcut to the full correct set.
5. Gate exact-value questions behind an estimation-first step (score a range or order-of-magnitude answer before revealing the precise question), rewarding number sense a solver does not need.
6. Interactive manipulatives — drag a point on a number line, place points to build a graph, move terms to balance an equation — graded on the resulting UI state, not a typed number: the one format a solver cannot hand back as a copyable string even when it can describe the answer.
7. In-app Socratic tutor chat as the primary hint-seeking path, so asking for help produces a graded dialogue instead of a screenshot-extractable string; score partly on coherence and specificity of the student’s own follow-up turns.
8. A short “explain it in your own words” field before an answer is accepted; use any AI-text signal only as a soft flag for human review, never an automated block, given detector bias against non-native writers and evadability by paraphrasing [7][8].
9. Ranked-mode time budgets short enough that a full external round-trip (photograph, OCR/solve, copy back) costs more than solving directly; untimed practice mode stays the openly low-friction, non-competitive surface.
10. Randomize numeric parameters server-side per student/attempt, so a screenshot, shared answer key, or cached web result does not transfer to another student’s identical item.
11. Weight scoring toward consistency across many small items (streaks, portfolios) rather than single high-value items, reducing the payoff of solving any one item via outside help.
12. Confidence self-rating (1-5) alongside each answer; miscalibrated high confidence with mismatched reasoning is a useful, non-punitive signal.
13. Split the integrity story by mode: practice makes no solver-resistance claim; ranked/leaderboard mode concentrates adaptive follow-ups, process-grading, and short timers, since that is the surface whose integrity matters to other users.
14. Route PhD-adjacent “produce a novel proof” items to asynchronous human or peer review rather than auto-grading — the one format that still resists both consumer solvers and frontier models [11][2][3], and the one format nobody can auto-grade at scale.

What we cannot prevent, stated plainly: any task fully specifiable as plain text or a single image with one verifiable final answer, submitted without a required process or follow-up, will be solved in seconds by tools already in wide student use — because the photo or live-vision channel never touches the app at all. Live multimodal assistants further erode time-pressure defenses, since a student can get spoken guidance while the screen stays visible instead of round-tripping a static screenshot. No detector-based gate is safe to use punitively. None of this is fixable by engineering; it can only be made less rewarding (not the scored surface) or more costly per point (adaptive design). A claim to the contrary is marketing, not fact.

Open questions for the project owner

1. Should ranked/leaderboard modes enforce a hard per-item time ceiling shorter than a typical photo-and-solve round trip — and what accommodation exists for students who genuinely need more time?
2. How much roadmap goes to an in-house Socratic tutor (own LLM cost, moderation, latency) versus adaptive-item and process-grading design with no generative component at all?
3. Should the product ever use AI-text-similarity signals in a bilingual EN/ES context, given documented detector bias against non-native writers — or is that class of check ruled out by policy?
4. For PhD-level content, is manual/peer review of open-ended proofs in scope, or does the top tier stay confined to auto-gradable formats (proof-critique, error-spotting) even if that caps how “PhD” it can be?
5. Should practice mode explicitly permit external tool use as a stated design choice, reframing the integrity narrative around ranked mode and mastery-over-time rather than implying practice answers are solver-resistant when they structurally cannot be?