Confirming a mailbox needs two network operations. A DNS query to find the domain's mail servers, and an SMTP conversation with one of them to ask about the recipient. Both return live answers from systems that change hourly, and neither can be recalled from training data.
A model asked whether john.smith@acme.com is valid will produce a confident-sounding response either way. It is pattern-matching against text, and a plausible-looking address at a real company is exactly the input that produces a confident wrong answer.
Code interpreter and browsing features do not close the gap. Sandboxed environments block outbound port 25 as a matter of course, since that is precisely the capability an abuse-prevention policy exists to remove.
Where a model genuinely helps is the surrounding work. Spotting that a column of addresses contains three different formats, drafting the regex for a first-pass syntax filter, writing the script that calls a verification API, explaining what a 550 in a bounce log means. All useful, none of it verification.
The same reasoning covers any tool without network access. A spreadsheet formula, a browser extension operating on text alone, an offline validator: each can check the shape of an address, and the shape was never the hard part.
A two-minute test with five addresses
You can see the limit for yourself. Pick five addresses where you know the truth. Use your own working address, and a colleague's. Add a mailbox at your company that was deleted last year and a made-up name at your own domain. Finish with an address at a domain that doesn't exist.
Paste them into a chatbot and ask which are valid. Watch what it goes on. It will likely approve the first four, because they all look like normal addresses at a real company. You may even get an explanation of its reasoning in a confident tone. The deleted mailbox and the made-up one are the tell, since nothing in the text separates them from the real ones.
Now run the same five through a real verifier. The deleted mailbox comes back invalid, because your mail server said so. Only a tool that asked the server can know that.
Put the model on the right side of the job
A model is good at the work around verification. Ask it to write a script that reads contacts.csv and sends each address in the work_email column to a verification API. Have it write the result into a new column. You'll have working code in a minute, and the check itself still happens where it should.
It's also handy for cleanup before upload. It can spot that one column mixes three date formats, or that some rows hold two addresses split by a semicolon.
Be careful with what you paste, though. A customer list is personal data. Dropping 5,000 names and addresses into a chat window shares them with another company, and your privacy terms may not allow that. Ask for the script, run it on your own machine, and keep the list out of the chat.
When a verification vendor advertises AI, ask one thing: does the verdict come from a server's reply, and can you see that reply?
Related questions
Can AI improve email verification at all?
At the edges. Pattern analysis can flag likely gibberish or guessed addresses. The decisive check is still a network conversation with the receiving server.
Why does ChatGPT sound so confident about addresses?
It is completing text rather than checking a fact. A plausible address at a real company is the input most likely to produce a confident wrong answer.
Can ChatGPT help with list cleaning?
Yes, for the surrounding work: normalizing formats, writing the integration code, and explaining bounce codes.
What about verification services that advertise AI?
Those run real SMTP checks and use models for scoring and pattern work on top. The protocol conversation is still doing the verification.