What does not count as evidence of AI
What does not count as evidence of AI
A project may present many artifacts that look like proof of a working AI component — but on closer inspection do not actually demonstrate one. This page lists the most common anti-patterns: signals that claim AI capability without providing verifiable evidence. It is the companion to What counts as evidence of an AI component.
In short: marketing copy, empty repositories, metrics without methodology, API wrappers with no proprietary logic, and "AI" in a token name are not evidence of a working AI component.
Why recognizing false signals matters
The boundary between "evidence" and "non-evidence" is the foundation of DecAIHub's approach to AI-washing. Recognizing the patterns below is not an accusation against the project — it means the specific artifact does not move the needle on verifiable AI reality, and should not be mistaken for Tier-1 evidence.
Common anti-patterns
"We use AI" in marketing copy
Saying "AI-powered," "built with machine learning," or "neural network-driven" on a landing page or pitch deck is a positioning statement, not evidence. Without Tier-1 artifacts (code, documentation, demos), these words carry no verifiable weight.
DecAIHub does not penalize projects for using the word — but it does not treat it as evidence either. The presence of "AI" in marketing raises the bar: if a project claims AI prominently, the absence of supporting artifacts becomes more conspicuous.
Empty or unrelated repository
A GitHub repository that exists but contains no AI/ML code — or contains only a README, boilerplate, or unrelated scripts — is a red flag. A meaningful repository has commits, releases, tests, and code that aligns with the claims in the whitepaper or documentation.
Common patterns: a repository created recently with a single initial commit; a fork of a popular framework with no modifications; a collection of configuration files with no model code; a repository with no commits in over a year while the project continues to claim active AI development.
Metrics without methodology
Claiming "95% accuracy" or "state-of-the-art performance" means nothing without context: what was measured, on what data, when, and with what version of the model. Metrics without methodology are not Tier-1 evidence.
"Our model achieves 0.92 F1 on the XYZ benchmark, evaluated on version 2.1 in January 2026" is a metric. "We achieve 92% accuracy" is a claim. DecAIHub records claims as noted statements but does not set the AI_3 flag based on them.
API wrapper with no proprietary logic
Calling the OpenAI, Anthropic, or another third-party API and returning the result is not evidence of the project's own AI. An API wrapper may be a valid product, but it does not make the project an AI project in the sense that DecAIHub evaluates. The AI belongs to the provider, not to the project.
If a project applies its own fine-tuning, combines third-party output with proprietary models, or adds measurable post-processing, the picture changes. The key question is whether the project adds verifiable AI-level value beyond what the raw API provides. If the entire "AI component" is a single API call with a system prompt, the AI reality belongs to the provider.
"AI" in token name or branding only
If a project's only connection to AI is the word "AI" in its token ticker, domain name, or branding — with no technical implementation behind it — this is positioning, not evidence.
A token ticker is a label, not a feature. DecAIHub treats naming as a neutral signal — it neither contributes to nor detracts from the evidence assessment. What matters is whether artifacts exist behind the name.
Partnership with an AI company
A partnership announcement (e.g., "partnered with NVIDIA" or "integrated with OpenAI") is Tier-2 at best. It shows association, not a working AI component inside the project. Partnerships do not substitute for code, documentation, or demos.
Partnerships range from deep technical collaborations (joint research, co-developed models) to loose affiliations (accelerator membership, being listed in a partner directory). Most announced partnerships in the crypto-AI space fall toward the looser end. Only partnerships with verifiable technical output contribute to the evidence picture, and even then as Tier-2 context rather than Tier-1 proof.
"AI roadmap" or AI intent without implementation
A roadmap item saying "AI features coming in Q3" shows intent (AI_0), not a working AI component. Intent is recorded as a single flag but does not, on its own, indicate a working AI component. The ai_score formula caps at 3, so AI_0 alone contributes minimally.
Plans change, get delayed, or get cancelled. DecAIHub records official AI intent through the AI_0 flag to distinguish projects that have declared AI ambitions from those that have not. But the flag is the floor, not the ceiling — it contributes ai_score = 1 at most, and only when the intent comes from a Tier-1 source.
Demos with no verifiable AI backend
A polished frontend that claims to show AI results but provides no way to verify that an AI model is running behind it — no API documentation, no model version, no reproducible outputs — is not sufficient evidence.
A verifiable demo allows different inputs and produces contextually varied outputs. It may expose an API endpoint, display model version information, or offer a sandbox mode. A showcase, by contrast, follows a scripted path or produces identical outputs regardless of input. DecAIHub looks for the former; the latter is noted but does not contribute to evidence flags beyond AI_0.
"Fine-tuned model" without specifics
Claiming to have a fine-tuned model is not evidence unless the project specifies: the base model, the dataset used for fine-tuning, the improvement over the base model, and quality metrics with methodology.
A credible fine-tuning claim names the base model, describes the training data, and shows measurable improvement on a defined task. Without these specifics, "we fine-tuned a model" is indistinguishable from "we used a model."
Anti-pattern reference table
| Anti-pattern | What it claims | What it actually proves |
|---|---|---|
| "AI-powered" on website | The project has AI | Nothing without Tier-1 artifacts |
| Empty GitHub repo | Open-source AI code | Repository exists; AI code does not |
| "95% accuracy" with no context | High-quality AI | A number was stated; validity is unknown |
| OpenAI API wrapper | The project "uses AI" | The project calls someone else's AI |
| "AI" in token name | AI-related project | A naming choice was made |
| Partnership with AI lab | AI capability | An association exists |
| "AI in Q3 roadmap" | Future AI component | Intent (AI_0) only |
| Polished demo, no backend evidence | Working AI | A frontend was built |
| "Fine-tuned LLM" with no details | Proprietary model | A claim was made |
How to distinguish false signals from weak-but-real evidence
Not every artifact that fails the Tier-1 standard is worthless. Some artifacts are genuinely weak evidence — insufficient on their own but meaningful in combination. The table below clarifies the boundary:
| Artifact | False signal (no evidence value) | Weak but real evidence |
|---|---|---|
| Repository | Empty or boilerplate only | Contains some AI code but lacks tests or releases |
| Metric | Number with no methodology | Number with partial methodology (missing one of four context questions) |
| Demo | Scripted showcase with fixed outputs | Interactive demo without API docs or model version |
| Partnership | Logo on a partner page | Joint technical output (co-authored paper, shared repo) |
| Documentation | Marketing whitepaper with no technical depth | Technical whitepaper that describes architecture but lacks version details |
Weak evidence may contribute to setting individual flags when corroborated by other artifacts. False signals do not contribute regardless of quantity — ten empty repositories are no stronger than one.
How DecAIHub records these cases
When no Tier-1 evidence for a working AI component is found, the project's ai_score remains at 0 or 1. If AI is actively promoted in marketing while evidence is absent, an AI-washing flag may be set on the card.
This is not an accusation — it is a record of the gap between claims and artifacts as of the verification date. When new evidence appears, the card is updated.
The ai_score sums the binary flags AI_0 through AI_4 (capped at 3). A project showing only anti-patterns from this page will typically have ai_score = 0 or ai_score = 1 (if AI_0 is set via an official roadmap), which directly limits its maximum DecAI Fit.
What this list alone does not prove
- Exhibiting one or more anti-patterns is not an accusation against the project. Many early-stage projects lack artifacts simply because the technology is not yet public.
- The absence of anti-patterns does not prove the AI is real. A project can avoid every red flag on this list and still have no working AI component.
- This list is a snapshot — anti-patterns may resolve as projects mature and publish new artifacts.
Frequently asked questions
Why does having "AI" in the token name not count as evidence?
A token name is a branding decision, not a technical artifact. Any project can choose a ticker that includes "AI" without building any technology. DecAIHub assesses evidence based on inspectable, verifiable artifacts — code, documentation, demos, metrics with methodology. A name does not meet any of these criteria. The token name is recorded in the project card as metadata but does not influence any of the five AI reality flags (AI_0 through AI_4).
Is a partnership with NVIDIA (or another AI company) evidence of AI?
No. A partnership announcement is classified as Tier-2 at best — it provides context but does not prove that a working AI component exists inside the project. Partnerships vary widely in depth: from deep co-engineering efforts to simple "ecosystem member" listings. Unless the partnership has produced verifiable technical output (a shared codebase, co-published research, a jointly maintained deployment), it remains an association rather than evidence. It may support a Tier-1 finding if one exists, but it cannot replace one.
What about projects that promise AI features in their roadmap?
A roadmap entry from an official Tier-1 source (the project's own blog, documentation, or governance proposal) is recorded through the AI_0 flag, which captures "declared AI intent." This sets ai_score to at least 1, distinguishing the project from those with no AI mention at all. However, a roadmap is a plan, not an implementation. The AI_0 flag is the lowest possible positive signal, and alone it limits DecAI Fit to a maximum of 3 out of 6 due to the ai_score ceiling rules. When the roadmap milestones are delivered and verifiable artifacts appear, the card is updated and higher flags may be set.
Can a project show multiple anti-patterns and still have a real AI component?
Yes. Anti-patterns indicate that specific artifacts do not constitute evidence — they do not prove the absence of technology. A project might have poor marketing practices and vague metric claims while simultaneously operating a legitimate AI system that is not well-documented publicly. DecAIHub evaluates what is verifiable at the time of review. If the underlying technology exists but is not externally inspectable, it cannot be scored — and the project is encouraged to publish the artifacts needed to update its card.
External references
- OECD AI Policy Observatory — AI classification framework — an international resource for understanding how AI systems are defined, classified, and governed across jurisdictions.
- Papers With Code — state-of-the-art benchmarks — a community resource that tracks AI benchmark results with linked code and methodology; useful for understanding what "metrics with methodology" looks like in practice.
- Stanford HAI — AI Index Report — an annual report tracking trends in AI research, development, and deployment; provides context for evaluating AI claims against the broader landscape.
No comments to display
No comments to display