Why the AI token label has become almost meaningless
In 2023, attaching the word 'AI' to a token could move its price. In 2024 and 2025, it still could. That is itself the problem. When a category grows faster than the underlying products can be built, marketing fills the gap, and the gap in crypto is enormous: a pitch deck, a website, a CEX listing, and a ticker are often all that stand between an idea and a billion dollars of attention.
This does not mean every AI-themed token is fraudulent. It means the label alone tells you nothing. FET, TAO, RENDER, VIRTUAL, VVV, and KITE all use the AI label, and they sit on a very wide spectrum, from networks that route real GPU compute and pay for inference to wrappers around public APIs with a token stapled on top. The reader's job is to learn which end of that spectrum a given project sits on before they risk money on it.
The rest of this article gives you a five-question filter, the on-chain signals worth checking, the GitHub and documentation tells that expose a hollow launch, and the news-context habits that keep you from buying the top of a narrative pump. None of this is financial advice. It is a way to read a project honestly before you decide whether it deserves your attention at all.
The core distinction: an AI project vs an AI-themed token
An AI project, in the strict sense used here, has a model, a pipeline that runs the model, and customers or users who get something from the pipeline. It does not have to be a frontier lab. A small open-source image classifier with ten paying API customers is an AI project. A token whose whitepaper mentions 'AI agents' on every page but ships no model, no API, and no users is not.
An AI-themed token is a cryptocurrency that uses AI vocabulary to attract demand. The token may have a real function inside the project's ecosystem, such as paying for inference or staking for governance, but the existence of the function is not the same as the existence of the product. A toll booth on an empty road still collects nothing.
This distinction matters because the failure modes are different. A real AI project can still fail as an investment if its token economics are bad, its team exits, or its model is out-competed. An AI-themed token can go up in price for months on hype and still have nothing of substance under the hood. The risk profile of the two is not the same, and treating them as the same is how beginners lose money.
Risks of buying AI-themed tokens without checking
The most common failure is paying a premium for narrative. When an 'AI agent' token rises 400% on a single announcement and the project has no shipped product, the buyers are usually late to a marketing push, not early to a real business. The price can hold for weeks and still leave the late buyer underwater once the team vests and starts selling.
Locked-team tokens make this worse. If the founding team's allocation unlocks months after launch while public buyers cannot sell at the same price, the structural setup rewards insiders and penalizes retail. This pattern is so common in AI-themed launches that it should be treated as a default assumption, not a surprise.
There is also the 'super-app' pitch risk. When a single token promises to be an LLM, an agent framework, a data marketplace, a compute layer, and a consumer app all at once, the project is usually pre-revenue on every axis. Shipping one of those is a multi-year job. Shipping five in parallel, with a token, on a 2025 roadmap, is almost always a story, not a product.
Finally, there is the API-wrapper trap. Some AI tokens are thin wrappers around OpenAI, Anthropic, or a public model. The wrapper adds little defensibility, the cost of the underlying API is the team's biggest expense, and the token does not capture that cost. If a competitor can copy the wrapper in a weekend, the moat is the marketing budget, not the technology.
The five-question filter you can apply in 60 seconds
Before you look at the chart, run the project through these five questions. If it cannot answer three of them clearly, treat the token as marketing.
1. Who uses the product, and how many?
Real AI projects have users. Not 'community members' or 'wallet holders', but identifiable people or agents paying for an output. The honest answer is a number, however small. 'We have 2,300 active inference jobs per day' is an answer. 'We are building for the future of decentralized AI' is not.
2. Who pays for the product, and in what currency?
If the project sells something, the customer pays in dollars, stablecoins, or the project's own token. The payment flow should be visible. A project that only 'rewards' users with emissions, and has no external revenue, is paying users with future buyers' money. That can work for a while. It cannot work forever.
3. What model is used, and is it documented?
Every serious AI project names the model, the training approach, or at minimum the architecture. A whitepaper that says 'we use state-of-the-art large language models' without naming one is not a technical document. Look for the actual model card, the dataset description, and the evaluation results.
4. Can the AI work be verified on-chain?
For decentralized AI, the proof should be visible on a block explorer. Subnet emissions on Bittensor, GPU hours on Render, inference transactions on a verifiable compute network: these are the on-chain equivalent of revenue. If a project claims to run AI but its chain shows nothing, the chain and the claim are not the same thing.
5. Is the code open, recent, and maintained?
Open source is not a requirement for a legitimate business, but in crypto it is the cheapest signal of seriousness. A GitHub with regular commits, real issues, and a maintained repo is a positive signal. A GitHub that was created the week of the token launch and has not been touched since is not.
On-chain verifiable metrics that actually mean something
Most crypto metrics are noise, but a few on-chain numbers map closely to real activity. For AI networks, these are the ones worth tracking.
Subnet emissions are the most direct read on Bittensor (TAO). Each subnet has miners producing outputs and validators scoring them. The emissions a subnet receives, and the trend of those emissions over weeks, are the closest thing the network has to a 'this subnet is being used' signal. Rising emissions on a specific subnet usually mean real demand for that subnet's output.
GPU hours are the equivalent on Render (RENDER) and similar compute networks. The number of jobs completed, the total compute delivered, and the diversity of providers tell you whether the network is actually routing work or just sitting idle. A network with a high token market cap and low GPU hours is priced for a future that has not arrived.
Active inference transactions are the signal on agent-focused networks like VIRTUAL and the newer VVV and KITE ecosystems. Count the number of paid inference calls per day, the number of unique agents, and the average spend per agent. If those numbers are flat while the token is up 5x, the buyers and the users are different groups of people.
For Fetch.ai (FET), the relevant on-chain signals are the number of autonomous agents registered, the volume of agent-to-agent transactions, and the activity on the agentverse. The token can rally on partnership announcements; the on-chain data tells you whether the partnerships produced anything.
GitHub and documentation tells that expose hollow launches
A project's repository is a public record of work. Reading it takes ten minutes and saves you a lot of money. Three tells matter most.
Commit history tells you when the work happened. Look at the last commit date, the frequency of commits over the past six months, and whether the contributors are full-time or sporadic. A repo with one big commit the week of the token launch and nothing since is a launch artifact, not a product.
Documentation depth tells you how serious the team is. Real AI projects publish model cards, API references, integration guides, and architecture diagrams. A whitepaper with glossy marketing language and no technical doc is a sign that the team has decided to spend its attention on buyers, not builders.
Issue and PR activity tells you whether the project has a community of users or just a community of speculators. Real users open issues, file bugs, and request features. Speculators open price posts. If the project's Discord and GitHub are 95% price talk and 5% technical talk, the user base is traders, not users.
Red flags that should end your evaluation immediately
No shipped product is the loudest flag. If the project has a token, a listing, and a roadmap, but no working software, you are buying a promise. Promises in crypto have a poor track record.
Locked-team tokens with short unlock cliffs are the structural flag. If the team holds 30% of supply and the cliff is six months, the launch is structured to enrich the team at retail's expense. The numbers are in the tokenomics doc. Read them.
An all-in-one 'super app' pitch is the ambition flag. Building one thing well is hard. Building a model, an agent framework, a data layer, a marketplace, and a consumer app, all with a token, is not a roadmap. It is a fundraising deck.
Anonymous teams with locked tokens are the trust flag. Anonymous founders are not automatically bad, but anonymous founders with locked tokens and a multi-year vesting schedule are asking you to trust people you cannot name with money you can lose.
API-wrapper products with no defensibility are the business flag. If a project is a thin layer over a public model, with no proprietary data, no fine-tuning, and no integration depth, the product can be cloned in days. The moat is the marketing.
Practical implications: how to use this filter before you act
Apply the five questions before you look at a price chart. If the project cannot answer three of them clearly, you do not have an investment decision to make. You have a marketing pitch to ignore. The chart will still be there if the project ships a product later.
Check the on-chain metrics weekly, not daily. Daily moves are noise. Weekly trends in subnet emissions, GPU hours, or inference transactions are signal. If the on-chain data is rising while the price is flat, the project is being used more than it is being talked about, which is a healthier state than the reverse.
Read the GitHub once a month. Five minutes on the commit graph and the issues page is enough to see whether the project is alive. If the last meaningful commit was three months ago and the last 'announcement' was a partnership, the project is in marketing mode.
Use news context with sentiment to spot pushes. When an AI token spikes on a single announcement, the question is whether the announcement is real adoption or a marketing push. Sentiment-aware news feeds, including Zipp's, score headlines as bullish, neutral, or bearish and rate their importance, so you can tell a partnership announcement with named counterparties from a vague 'strategic alignment' press release.
How to follow AI tokens the smart way
AI tokens move fast and so does the news around them, and most of that news is narrative, not substance. Tracking announcements, partnership claims, and unlock schedules manually is a losing game, because by the time you have read ten threads the next push has already started. Zippfeed surfaces AI-token headlines with sentiment scoring, bullish, neutral, or bearish, and an importance rating, so you can see which stories are actually moving the conversation and which are just moving price. Pair that feed with the five-question filter above, and you have a faster way to tell which AI tokens are doing real work and which are doing real marketing.
CoinDesk
