Valid Data Is Not a Real Person: The Human Data Illusion
A name, email and phone number prove a record exists, not that a person acted. Four cognitive biases explain why fake leads and synthetic records stay in the CRM.
Key findings
- A record with a name, email address and phone number proves that those details exist. It does not prove who submitted them, that the submitter owns them, or that anyone asked to be contacted.
- Four cognitive biases keep bad data in place: anthropomorphism, sunk cost, the illusion of progress and loss aversion. They stack in that order.
- Around 34% of applications to California’s community colleges were flagged as likely fraudulent. One district flagged 64%. Nearly $5.6 million in federal aid went to fake students in three months.
- JPMorgan paid $175 million for a customer base of 4.25 million. About 300,000 were real. The rest was a synthetic data set that passed due diligence.
- Valid no longer means genuine. Treat every record as a set of claims and ask what can be proved.
Summary assessment
The human data illusion is the tendency to treat a data record as a real person. A row holding a name, an email address, a phone number and a date of birth is read as an individual with intent. The reader supplies everything the record does not contain.
The assessment of this file is that the illusion is now the main weakness in lead generation, CRM data, list data and commercial identity data. The working question has changed. It used to be “is this data valid?” It is now “is there a person here at all?”
Source: high. Simon Delaney handles lead generation traffic and took part in detecting the third incident. Background in lead generation, identity, verification and software for data processing.
Evidence: published for the mechanism and the two public cases, in state reporting and court record. Stated for the third incident. Contact source for evidence. The trend is the source’s view of the traffic he handles. This file shows how it is done. How widespread it is has not been measured.
How a data record becomes a “person”
Place a spreadsheet row in front of an analyst: Mr John Smith, a mobile number, an email address, a date of birth, an IP address. The analyst pictures a man. Nobody describes a CRM as 48,000 rows containing values. They say there are 48,000 people in it.
In lead generation the same step happens faster. Name plus email plus phone number becomes “a prospect”. What the record establishes is narrower:
- A name, an email address and a phone number were submitted together.
- It does not establish who submitted them.
- It does not establish that the submitter owns those details.
- It does not establish that anyone asked to be contacted.
The prospect is constructed by the reader. The record only supplied the materials.
Four cognitive biases that keep fake leads in the CRM
The source identifies four psychological mechanisms. None is new. Each is well documented outside the data industry. Together they explain why bad data is accepted and then defended.
1. Anthropomorphism: seeing people in records
Anthropomorphism is the habit of attributing human qualities to things that are not human. It is why people see faces in clouds and plug sockets. Applied to data, it converts a record into a person with motives. This is the entry point. Without it the other three biases have nothing to protect.
2. Sunk cost: the data has been paid for
Once money, time or reputation has gone into a data source, evidence against it is resisted. A buyer who has spent £20,000 on leads and is told that 30% look fraudulent rarely deletes them. Observed responses include:
- “Are we sure?”
- “Can sales try them again?”
- “Maybe the source had a bad week.”
- “The rejection criteria are too aggressive.”
- “There may be a fault in the verification system.”
Accepting the finding means accepting that part of the investment was wasted.
3. The illusion of progress: the number went up
More leads, more records and a fuller pipeline feel like growth. A rising dashboard produces that feeling whether or not revenue follows. Volume is counted because it is visible. Conversions are the only figure that tests whether the records were real.
4. Loss aversion: deletion feels like destruction
Data inside a business starts to feel like an asset. Removing it feels like destroying something owned. Indicators include dead leads kept for years, very large legacy databases, disputes with suppliers over rejection rates, and continued contact long after the evidence says the records have no value.
How the four biases stack
They occur in sequence:
- This looks like a person.
- We paid for it.
- Look how much the database has grown.
- We cannot throw this away.
By the fourth step the data is defended by the people best placed to question it.
Evidence: three incidents
Incident 1: California Community Colleges
Status: corroborated by state reporting.
Around 34% of applications across California’s community college system were flagged as likely fraudulent. In the Los Rios district the figure reached 64% between January and March 2025. In those three months, nearly $5.6 million in federal student aid and more than $900,000 in state aid was paid to fake students. Methods included bots and stolen identities.
Reading: the application said “student”. The identity looked like a student. The system counted a student. At no point was there a person intending to study.
Incident 2: JPMorgan Chase and Frank
Status: corroborated by court record. The founder was convicted in March 2025 and sentenced to 85 months.
JPMorgan Chase acquired the student finance company Frank for $175 million in 2021. Most of the value sat in its claimed customer base of 4.25 million users. The company had about 300,000. A “user” had been defined to the bank as a record with four fields: first name, last name, email address and phone number. A synthetic data set was produced to match the claim, and a third-party vendor confirmed that it contained more than 4.25 million rows.
Reading: due diligence counted rows. Each row had the four fields that make a record look like a person. Nobody tested whether the people existed. The data scientist who built the synthetic set testified that a few calls would have exposed it.
Incident 3: lead supplier, identity withheld
Status: first-hand. The source took part in the detection.
A single supplier submitted high-value, supposedly high-intent leads into a network for roughly four years. Volume is put at hundreds of thousands, possibly millions. The supplier was recently caught. The source was directly involved in the detection.
The records were good. Names, emails and phone numbers were valid and passed every field check. Three things gave the supplier away:
- Volume and consistency. Too many leads, arriving too evenly.
- Submission metadata. A capture layer between the network and the end client recorded what the network could not see. The metadata was the same on every submission and the timings were mechanical.
- Breach exposure. The details in every record traced to breached data.
The account team did not act on the evidence and kept asking for more. Senior staff outside the account reviewed the same material and acted on it.
Reading: every field was real and no person was present. Validation could not have caught this. The metadata and the pattern across the batch did.
The same mechanism in AI
Mustafa Suleyman, chief executive of Microsoft AI, has warned about what he calls “Seemingly Conscious AI”: systems that show enough signs of a person that users fill in the rest and believe someone is there.
The mechanism is identical. A chatbot presents human signals and the user supplies an inner life. A record presents human fields and the reader supplies a person, a motive and intent. AI also lowers the cost of producing those fields at volume.
Where the illusion applies
- Lead generation, B2C and B2B. Records are paid for on submission, before intent is tested.
- CRM and list data. Size is reported as a strength.
- Commercial KYC and identity databases. The source reports reference files of hundreds of millions of records, a material share of them old or of doubtful origin, which holders will not reduce. See where commercial KYC data comes from.
- B2B content syndication. Buyers keep running campaigns after the leads fail to recall the content. See real B2B contacts, invented engagement.
- Fraud detection. Checks built to confirm validity pass real details used by the wrong party.
Assessment: valid does not mean genuine
Four assumptions no longer hold:
- Valid data is not genuine data. Details can be correct and still not come from their owner.
- Real data is not real activity. A real person’s record can be submitted by a bot.
- A real identity is not ownership. Whoever submitted an identity may not be the person it describes.
- An identifiable person is not intent. A fully identified individual with accurate details may never have made the enquiry.
The current default is that a record is a person unless proved otherwise. The recommended default is the reverse: a record is a set of claims. What can be proved?
The countermeasures follow below.
Countermeasures: Data that only looks human
Stolen and generated records pass every field check. Validation proves the details are real. It does not prove that a person submitted them.
- Start from unproven
A delivered batch is not a result. Until it is tested, each record is neither valid nor invalid. Receiving leads is not success. Do not count or report a record as a person before it has been checked.
- Validate every field
Check that the name is real, the email delivers and the number can be reached. This removes honest mistakes, because real people mistype. It does not remove fraud. Invalid data does not get paid, so fraudsters rarely send it.
- Read the submission metadata
Look at how each record arrived: device, browser, IP address, time taken and the gaps between submissions. Ask whether it looks like one person completing one form. The same metadata on every submission is one operator, not a crowd.
- Test the batch as a whole
People are messy and inconsistent, in their details and in how they submit them. Look across the batch for anything mechanical: steady volume, regular timing, uniform records, breach signals that repeat. Check that the demographic fits the value of the product.
- Cross-check against an independent source
Borrow from identity verification and KYC. Check whether elements such as date of birth match what a known high-quality source holds for that person.
- Capture the submission yourself
Put your own capture between the source and your system. A lead can be passed on without the metadata behind it. In the supplier case in this file, a capture layer between the network and the end client recorded bot metadata that had been masked from the network.
- Pay only for records that pass
Agree before buying that payment is due only on records you find valid, and give the supplier a reason for every rejection. Set the trigger in advance. For high-value leads, one bad record should prompt action. For low-cost co-sponsored leads, set a percentage.
- Decide from the top
Whoever is responsible for the campaign, and for the company or client it serves, owns the decision to cut a source. Senior people need sight of the evidence. Act on what you have. Often there will be no more.
What does not work
- Field validation aloneIt proves the details exist. Stolen data is valid data.
- Fraud detection as a wallIt is a balance between blocking everything and letting good traffic in. Fraudsters work to look like good traffic, and some get through.
- CAPTCHABots complete it.
- Double opt-inBots complete that too, and SIM swaps defeat confirmation by phone.
- Consent certificatesA record of where and when a submission happened. Not evidence of who made it.
- AI validationAI checks syntax and spots patterns. It cannot confirm that a person is real.
Limits
No single check fixes this. Personal details, first-party data and submission metadata are already collected in almost every form of online data capture. Almost nobody reads them together. Reading them together, record by record in real time and across the batch for patterns, is the only method that works.
Example checks
- name
- real
- deliverable
- phone
- reachable
- address
- exists
- metadata
- same on every submission
- timing
- mechanical
- breach
- whole batch exposed
- errors
- none
Example checks are illustrative. They are not real records.
Operational resources
Relevant services and organisations that can help implement these countermeasures.
Provero was founded by Simon Delaney, who produces The Verification Files.
- Provero: Form verification
Runs checks behind a form as it is submitted: contact details, connection signals and submission metadata. Valid contact details alone do not prove that a real person completed the form.
Supports: 02 Validate every field · 03 Read the submission metadata
A resource is one way to carry out a countermeasure. Listing is not an endorsement. Some countermeasures may only be able to be applied via a resource. How resources are chosen
Frequently asked questions
What is the human data illusion?
The human data illusion is the tendency to treat a data record as a real person. A row containing a name, an email address, a phone number and a date of birth looks like someone, so the reader assumes a real individual with real intent stands behind it. The record proves only that those details exist together in one place.
Does valid data mean a lead is genuine?
No. Validation confirms that an email address delivers, a phone number connects and an address exists. It does not confirm that the owner of those details submitted them or wanted contact. Stolen and recycled identities pass validation because the details belong to real people who took no part.
Why do businesses keep fake or low-quality leads?
Four cognitive biases work together. Anthropomorphism turns records into people. Sunk cost makes paid-for data hard to write off. The illusion of progress makes a growing database feel like growth. Loss aversion makes deletion feel like destroying an asset. The result is that evidence of fraud is met with requests for more evidence.
How did synthetic customer data pass due diligence at JPMorgan?
In the 2021 acquisition of the student finance company Frank, the seller claimed 4.25 million users and supplied a synthetic data set with that many rows to a third-party vendor, which confirmed the row count. The company had about 300,000 users. The check counted records. It did not test whether the people in them existed.
How should a business treat a new lead or customer record?
As a record making claims, not as a person. Each claim needs its own evidence: that the details are valid, that they belong together, that their owner submitted them, and that the submission reflects real intent. A record that passes only the first test is unproven.
How do you detect fake leads that pass validation?
Look past the fields. Check the submission metadata: device, browser, IP address and timing. Then check the batch as a whole for mechanical patterns such as identical metadata, regular timing and breach exposure across every record. Real people are messy and inconsistent. Automated submissions are not.
Do CAPTCHA and double opt-in stop fake leads?
No. Bots complete CAPTCHA and double opt-in, and SIM swaps defeat confirmation by phone. A consent certificate records where and when a submission happened. It does not show who made it.
Exhibits: sources and further reading
- AUS Department of Justice: Startup CEO sentenced to 85 months in prison for $175 million fraud (Frank / JPMorgan Chase) ↗
- BCalMatters: Fake student aid, California colleges detect more fraudsters stealing millions (April 2025) ↗
- CCalMatters via LAist: California community colleges crack down on fake students stealing financial aid ↗
- DMustafa Suleyman: Seemingly Conscious AI is Coming (August 2025) ↗