Read before relying on it
What PhantomAI cannot do
Most AI products describe only their capabilities. That leaves you to discover the limits yourself, usually on something that matters. So here is the specific version.
Two kinds of limitation are listed separately: things that are true of every AI language model, and things that are true of PhantomAI specifically as a product. Both matter, for different reasons.
Limits that come from the AI itself
PhantomAI runs open-weight models (openai/gpt-oss-120b and openai/gpt-oss-20b) served through Groq. These limits apply to it exactly as they apply to every other general-purpose assistant.
It can state things that are false
PhantomAI generates text by predicting plausible continuations, not by looking facts up. It will occasionally produce confident, well-written statements that are wrong. This is inherent to how language models work, not a bug we can patch.
It will invent citations and sources if you ask for them
Do not ask PhantomAI for academic references, case law, or reading lists. It produces correctly formatted references to papers that do not exist. Use Google Scholar, PubMed, or your library, then paste what you find into PhantomAI.
It is unreliable at arithmetic and counting
Calculations are predicted rather than computed. Ask it to write a formula or a script instead of asking for the number, and check the result yourself.
It has no reliable sense of its own uncertainty
The confident tone is a writing style learned from training data. It reads the same whether the answer is solid or invented, so tone tells you nothing about accuracy.
It may change its answer if you push back
Asking “are you sure?” tests how agreeable the model is, not whether it was right. It will sometimes abandon a correct answer. Verify against a source instead.
Its knowledge has a cutoff, and search is not perfect
Without searching, it does not know about recent events. When it does search, it can still misread or over-summarise a page. Open the source before relying on a claim.
Long conversations lose earlier detail
There is a limit to how much conversation fits in a single request. In long threads, earlier instructions can drop out. Start a fresh chat for a new topic and restate anything critical.
Answers are not perfectly consistent
The same question can produce different answers, because token selection involves randomness. Re-asking is not a way to verify something.
Each of these is explained in more depth, with tests you can run yourself, in limitations of AI assistants.
Limits of PhantomAI as a product
These are features that genuinely do not exist. If you need one of them, it is better that you know now.
It cannot run or test your code
PhantomAI has no code execution environment. When it says code works, that is a prediction. Run it yourself, including edge cases.
It cannot act on your behalf
It cannot send emails, book anything, make purchases, log into services, or change settings in other apps. It produces text and files; you take the actions.
It cannot see images you have not given it
It only works with what you paste, type, or upload in the conversation. It has no access to your device, your files, or your other accounts.
There is no team or collaboration layer
Accounts are individual. There are no shared workspaces, no shared conversation libraries, no admin console, and no centralised billing. If you need those, PhantomAI is not the right tool today.
There is no mobile app
PhantomAI is a web app. It is responsive and installable as a PWA on most phones, but there is no native iOS or Android app in an app store.
There is no API
You cannot call PhantomAI programmatically from your own software. There are no API keys to issue.
Usage is capped by a daily credit allowance
Every account gets a daily credit balance that refreshes. Heavy use in a single day can exhaust it, and you will need to wait for the reset.
Conversation history depends on your account
Chats are stored against your account. If you lose access to your login, you lose access to the history. There is no export-everything button yet.
What it is not a substitute for
Medical
Not a doctor, not a diagnosis, not a treatment plan. Useful for understanding terminology and preparing questions for a clinician.
Legal
Not a lawyer, and conversations carry no privilege. Law is jurisdiction- and date-specific in ways generated text does not reliably capture.
Financial & tax
Not advice. Rules vary by country and change yearly. Use it to learn vocabulary, not to make decisions.
Mental health & crisis
Not a therapist and not a crisis service. If you are in danger or in crisis, contact local emergency services or a crisis line.
Safety-critical work
Do not rely on it for electrical, structural, chemical, or dosage instructions. Consult a qualified professional and an authoritative source.
Academic submission
Submitting generated text as your own writing is misconduct at most institutions. Check your own policy — they vary by course.
Where it genuinely works well
A limitations page that only lists failures is not honest either. There is a clear pattern to when these tools are reliable: reliability tracks how little the model has to invent. Give it the material and ask for a transformation.
- Reshaping text you provide — summarising, restructuring, changing tone, turning notes into a table.
- Answering questions about a document or URL you supply, because the facts come from your source rather than its memory.
- Explaining well-documented concepts at whatever level you ask for, repeatedly, without impatience.
- Generating options you will filter yourself — names, counterarguments, approaches, test cases.
- First drafts of conventional, structured work you intend to review anyway.
- Critiquing your own explanation or your own code, which is where a tireless second reader is genuinely valuable.
Security and privacy, stated plainly
PhantomAI is built and run by one person. There is no dedicated security team, and no formal compliance certifications — no SOC 2, no ISO 27001, no HIPAA coverage. Claiming otherwise would be dishonest.
Your conversations are stored so history persists, and messages are sent to a third-party inference provider (Groq) to generate responses. Conversations are not end-to-end encrypted, because a model has to read your text to answer it. We do not train models on your conversations.
Practical advice: do not paste passwords, API keys, payment details, other people's personal data, or material your employer has not cleared. See AI privacy basics and our privacy policy.
Common questions
Why does PhantomAI publish a limitations page at all?
Because “what can it not do” is one of the first questions a reasonable person asks, and because overstating what an AI assistant can do is the most common way these products mislead people. Everything on this page is checkable by using the product.
Does PhantomAI ever give wrong answers?
Yes. It generates plausible text rather than retrieving verified facts, so it can produce confident, fluent statements that are false. Verify anything that matters, especially numbers, dates, citations, and technical specifics.
Can PhantomAI run the code it writes?
No. There is no execution environment. Any claim that code works is a prediction, not a test result. Run it yourself.
Can I use PhantomAI for medical, legal, or financial decisions?
No. It is not a substitute for a qualified professional in any of those areas. It can help you understand terminology and prepare better questions, which is genuinely useful, but the decision needs a professional.
Is there a team plan or an API?
No. Accounts are individual, and there is no API, no admin console, and no shared workspaces today. PhantomAI is a solo-built project and we would rather say that than advertise features that do not exist.
Found something on this page that is wrong?
If PhantomAI does something this page says it cannot do, or fails at something we have claimed it handles well, tell us and we'll correct it. Email phantomai2009@gmail.com.