How to Review an AI-Generated Product Concept Without Fooling Yourself

Save the prompt, model or system if known, date, output, images, and revision. Do not begin by editing away contradictions. Highlight every number, material, standard, safety statement, performance claim, market fact, patent implication, supplier assumption, and environmental claim. Label each proposed, sourced, observed, verified, or specialist-approved. Most generated statements should begin as proposed.

The NIST AI Risk Management Framework encourages explicit management of AI risks and trustworthiness. For product work, a simple control is provenance: who or what produced the statement, which source supports it, what test applies, and which decision depends on it. Fluent language is not a source. A citation must actually support the specific claim and applicable context.

Ask whether the output names a real user situation, current alternative, measurable outcome, and failure consequence. Generated concepts often jump from a broad problem to a feature bundle. Rewrite adjectives as measurable requirements with conditions, tolerances, and verification. Separate user need from proposed solution so an attractive feature can be removed without losing the problem.

Look for conflicts: cheap and premium materials, waterproof and easily opened, tiny and long-runtime, universal and tightly fitted, automated and privacy-preserving. Conflicts are not automatically fatal, but they need priority and evidence. Record the tradeoff instead of allowing different sections to promise incompatible outcomes.

Trace every part, interface, motion, load, wire, seal, fastener, vent, sensor view, battery path, user hand, tool path, and assembly sequence. Renders can hide interference, unsupported parts, impossible access, missing thickness, disconnected components, and inconsistent scale. Create section views and simple interface diagrams. Give critical dimensions and forces an owner and method.

Use calculation and bench prototypes appropriate to the consequence. A plausible geometry is not engineering validation. Check center of mass, stability, structural path, thermal path, tolerance accumulation, power budget, communications, failure state, and maintenance access where relevant. Ask a qualified engineer to review consequential mechanisms and document assumptions they could not verify.

Demand a cost boundary. Does the number include only visible components, or material, process, assembly, tooling, yield, inspection, packaging, freight, duty, warranty, returns, and channel margin? Is it tied to volume, origin, revision, currency, and date? Generated BOMs commonly omit fasteners, electronics support parts, test, fixtures, finishing, labor, and minimum order.

Challenge process compatibility, tolerance, wall strategy, tooling access, finish, material grade, joining, and inspection. A part can be printable but not production-ready. Use the first-pass cost model and process picker, obtain supplier feedback, and retain low, base, and high ranges. Do not present an AI estimate as a quote.

List intended use, users, markets, energy and material hazards, foreseeable misuse, and claims. Use official starting points such as the CPSC Regulatory Robot or FDA Device Advice where relevant, then escalate to qualified review. An AI system cannot certify the product, select every applicable standard, or know unprovided details. Remove certification marks and compliant language until evidence and authorization exist.

Review marketing claims under FTC advertising guidance. Technical-looking numbers can still be misleading if the method, conditions, or basis are absent. Check health, safety, performance, origin, environmental, and comparison claims especially carefully. Maintain a prohibited-claims list until the evidence pack supports specific language.

Run a prior-art discovery pass on the problem, functions, mechanisms, and classifications. Do not treat an original-looking render as legal novelty. Record close documents and how they change the brief. If the output cites sources, open every one, confirm publisher and date, and check whether it supports the exact statement. Watch for fabricated titles, outdated pages, summaries that overreach, and circular citations.

Separate research from legal conclusions. A quick search cannot establish patentability or freedom to operate. If commercial decisions depend on intellectual property, give qualified counsel the current concept, intended markets, search log, closest documents, and questions. Repeat the search after a material mechanism change.

Score each assumption by consequence if wrong, likelihood of error, current confidence, and cost to test. Select the few uncertainties that can reverse the project. For each, define evidence, method, owner, budget, decision date, and stop condition. Do not build the highest-fidelity prototype by default; build the cheapest credible test of the riskiest assumption.

Finish with a red-team record: claims removed, requirements rewritten, conflicts found, sources verified, hazards escalated, prior art captured, cost ranges corrected, lifecycle gaps, prototype tasks, specialist gates, and decision. ConjureAnything is valuable when it accelerates this loop. The output becomes stronger because reviewers are invited to disprove it.