Can I Paste This Into AI?

What this tool does

CM8-330 answers the question people ask themselves several times a day and almost never answer properly: can I paste this into a chat assistant? You put the text in, and the tool reads it the way a careful data protection officer would — looking for names, contact details, addresses, employee and customer identifiers, bank and card details, national and tax numbers, credentials, confidentiality markings and subject matter that carries special legal protection. It scores what it finds, gives you a verdict, shows you exactly where each item sits in your text, and hands you a redacted version you can copy instead.

It runs entirely inside this one file. There is no account, no upload and no network request of any kind — which is the only arrangement that makes sense for a tool whose whole purpose is to be shown your most sensitive text.

Why this matters more than it used to

Assistants are now genuinely useful for ordinary work: rewriting a difficult email, explaining an error log, summarising a contract, drafting a policy. The friction that used to protect organisations — the fact that sharing data was awkward — has gone. Pasting is free, instant, and invisible to everybody except the person doing it.

The result is that the most common data incident in most organisations is no longer a hacked server. It is a competent, well-intentioned employee pasting a spreadsheet extract into a browser tab at half past four on a Friday because they wanted help with a formula. Nothing about that person is careless. They simply had no practical way to find out what was in the text before it left.

This tool is that practical way. It is deliberately not a blocker, a policy engine or a monitoring product — it reports to the person doing the pasting, tells nobody else, and leaves the decision where it belongs. Tools that inform people are used; tools that police them are worked around.

How to use it

• Paste into the box. Scanning is immediate and continuous as you type; there is no send button, because there is nowhere to send it to.
• Read the verdict, then the marked-up copy. The verdict is a summary; the marked-up copy is the evidence. Highlight colour is sensitivity — red is critical, blue is background detail.
• Set the destination. The same text carries different exposure depending on where it is going, and the score adjusts accordingly.
• Tune the findings. Untick any type you want left alone. A supplier's public company name usually does not need redacting; the buyer's home phone number does.
• Copy the safe version — not the original — and check that it still contains enough for the assistant to be useful. If it does not, that is itself a finding: the task may not be one to delegate.
• Log the check if you want a record that checking happens. The log keeps the date, the score and the counts. It never keeps the text.

What it looks for

Sixteen detector families run over every scan, each carrying a sensitivity from 1 to 5.

• Family — Sens. — How it is found
• Credentials, keys and tokens — 5 — Private key blocks, JSON web tokens, vendor key prefixes, and any value written after the words password, secret, api key, bearer or connection string
• Payment card numbers — 5 — 13 to 19 digits that pass the Luhn checksum
• Bank account numbers (IBAN) — 5 — Country code, check digits and account body that satisfy the ISO 13616 modulo-97 test
• National, tax and identity numbers — 5 — A value written after a recognised label — national insurance, social security, tax ID, passport, driving licence and their common variants
• Special-category subject matter — 5 — Health, disciplinary, grievance, immigration, union, religious, ethnic and criminal-record vocabulary
• Other bank and payment details — 4 — Sort codes, routing and SWIFT/BIC codes, account numbers and card security codes, by label
• Dates of birth — 4 — A date written after a date-of-birth label
• Postal addresses — 4 — A number followed by a street-type word, and anything written after an address or postcode label
• Unlabelled high-entropy strings — 4 — Long mixed-character strings whose Shannon entropy exceeds 3.5 bits per character
• Email addresses — 3 — Standard address shape
• Telephone numbers — 3 — 7 to 15 digits with an international prefix, brackets, separators, or a preceding phone label
• Personal names — 3 — Honorific followed by a name, a name after a labelling word, or — optionally — any capitalised pair not on the exclusion list
• Employee, customer and record identifiers — 3 — A code written after employee, customer, account, invoice, case, ticket and similar labels
• Confidentiality markings — 3 — Confidential, internal use only, commercial in confidence, without prejudice, NDA and similar
• Network addresses and system paths — 2 — Web addresses, IPv4 addresses, MAC addresses, internal host suffixes and file paths
• Currency amounts — 2 — Currency symbols and codes, and figures near salary, margin, settlement and pricing words

Where two detectors claim the same span of text — a card number is also a long run of digits — the more sensitive claim wins, then the longer one. Each character of your text is counted once, never twice.

Confidence, and why it is shown

Every finding carries one of three confidence levels, and the difference between them is the difference between a tool you trust and one you learn to ignore.

• Verified — a checksum passed or a fixed vendor prefix matched. A Luhn-valid card number or a modulo-97-valid IBAN is essentially never an accident. Treat these as certain.
• High — the shape is unambiguous, or your own text labelled it. An email address is an email address; a number written after "national insurance number" is what it says it is.
• Possible — a heuristic. Unlabelled personal names, street addresses and high-entropy strings live here. These will produce false positives, and saying so plainly is the point: a tool that claimed certainty about "Fleet Street" being a person would deserve to be ignored within a week.

How the score is built

Each detected type contributes according to its sensitivity and, more weakly, its volume. Volume is deliberately damped: twenty email addresses are worse than one, but nowhere near twenty times worse, whereas a single verified credential outranks a page of names. Anything else would let a long, dull document out-score a short, catastrophic one.

Two overrides sit above the arithmetic. A checksum-verified credential or account number forces Do not paste outright, because that is not a matter of degree. Special-category subject matter alongside an identified person forces at least High risk, because brevity does not make it safe.

Where the text is going

The same paragraph carries different exposure depending on its destination, so the tool asks.

• Public assistant. Assume the text may be retained, may be reviewed by a human for quality, and may be used to train future models unless the provider has committed otherwise in writing. No reduction is applied.
• Business tier with a no-training agreement. A real reduction — 28% — but not an exemption. The data still leaves your control, the commitment binds the provider rather than an attacker, and a contract has never once prevented a breach; it only decides who is liable afterwards.
• Self-hosted or private deployment. A 55% reduction. The text stays inside your estate, so what remains is mostly about who can read the prompt logs, and how long they are kept.

The reductions apply to the score, never to the overrides. A private model is not a reason to paste a live credential into a prompt.

The safe version

Every finding you leave ticked is replaced in the safe version. Four styles are available in Settings. Type labels are the default and usually the best: an assistant that can see [PERSON] emailed [EMAIL] about invoice [REF] still understands the structure of your problem, and answers it properly. A solid block or a row of asterisks tells it nothing, and the reply is worse for it.

Read the safe version before you send it. If removing the sensitive parts has left something the assistant cannot possibly work with, you have learned something important — the task depends on the confidential detail, and needs a different approach rather than a cleverer redaction.

What it cannot find

This is a pattern matcher. It does not know what your text means, and there are whole categories of sensitive information it will never see:

• Identification by context. "The only night-shift supervisor on line 3" names a person to everyone at the site, and matches nothing.
• Commercially sensitive facts. An unannounced acquisition, a site closure, a pricing change or a lost tender is plain English with no shape at all.
• Project and product codenames, internal system names, and anything else meaningful only inside your organisation.
• Structure and volume. A thousand rows of ordinary-looking data may constitute a disclosure that any single row would not.
• Anything in an image or an attachment. Only the text in the box is examined.

Equally, it over-reports. Turn the unlabelled-name detector off when your text is dense with proper nouns and the real findings are drowning.

Why the text is never stored

The box you paste into is the only part of this tool that is deliberately excluded from saving. It is not written to browser storage, it is not included in the backup file, and it is not reproduced in the printed report — where the report shows masked examples only. Closing the tab destroys it.

This is a design decision, not an oversight. A privacy tool that quietly accumulated a local archive of every sensitive thing anybody checked would be the largest single concentration of risk in the building. The log records that a check happened and how it scored; it never records what was checked.

The formulas

The volume factor is capped at four doublings, so a type found sixteen times contributes 2.4 times what it would once, not sixteen times. That cap is a judgement, and it is the number to change first if the scores this tool produces do not match how your organisation actually weighs bulk disclosure.

The final curve saturates instead of truncating. A raw total of 45 scores 63, 90 scores 86, and 300 scores 99.9 — the score approaches 100 without ever reaching it. That matters for a practical reason: with a hard cap every seriously loaded document scored exactly 100, which made the destination setting look broken and made a bad text indistinguishable from a catastrophic one. Read the raw total in the findings table when you want the uncompressed figure.

FAQ

Does anything I paste leave my computer? No. The file contains no network code of any kind — no fetch, no request object, no embedded frame — and this is verified mechanically before release. You can also work offline: disconnect and the tool behaves identically.

It flagged a place name as a person. That is the unlabelled-name detector, which is honestly reported as "possible". Untick the type, or switch the detector off in Settings.

It missed something obviously sensitive. Almost certainly something identifiable by meaning rather than shape. Read the limits section — and treat a clear verdict as "nothing matched", never as "safe to send".

Why does the score fall when I choose a self-hosted model? Because the exposure genuinely is lower — the text does not leave your estate. The overrides still apply: a live credential is a bad thing to put in any prompt, anywhere.

Can I use this to prove compliance? It gives you a defensible record that checking took place, which is worth having. It is not an assessment of lawfulness, and it does not know your policies, your contracts or your jurisdiction.

Should the safe version be pasted as-is? Read it first. Redaction changes meaning, and occasionally changes it enough that the question is no longer the one you meant to ask.

Saving your work

Settings, the destination, your redaction choices and the check log are written to this browser's local storage. The text in the paste box is not, by design.

Treat Export .json as the real save — one file holding the settings and the log, which Import .json restores anywhere. Export CSV gives you the findings table for a spreadsheet. Reset asks twice, then erases everything this tool has stored. There is no undo.

Accuracy & disclaimer

This tool finds patterns; it does not understand your text. It will miss sensitive information that has no recognisable shape, and it will occasionally flag ordinary words that look like identifiers. A clear verdict means nothing matched the patterns it knows — not that the text is safe to share.

Nothing here is legal advice, and no score is an approval. Where a disclosure would breach a contract, a data protection law or a duty of confidence, the decision belongs to the people accountable for it, on the facts of the case. Use this to make that decision better informed, not to make it for you.