You have been asked to approve a 40% uplift for someone whose job title changed but whose team did not. The benchmark data you will be shown to justify it is real. It is also, most likely, measuring a market you do not hire in.
The AI/ML pay premium over traditional data science is one of the most confidently quoted numbers in hiring right now, and one of the least stable. Depending on which published source you open this month, the gap is 54%, or 15%, or roughly nothing at all. All three are correctly calculated. They are measuring different things.
Here is what the current data actually says, and what to do with it.
The 30 to 50 percent figure is real, on exactly one measure
Self-reported total compensation from levels.fyi, updated August 2026, United States:
| Role | 25th | Median | 75th | 90th |
|---|---|---|---|---|
| Data scientist | $131,000 | $180,000 | $250,000 | $350,000 |
| ML/AI software engineer | $166,000 | $245,000 | $364,000 | $509,300 |
| Machine learning engineer | $195,250 | $278,125 | $377,500 | $495,000 |
On median total compensation, the machine learning engineer title carries a 54% premium over data scientist. Take the broader ML/AI cut and it is 36%. So the headline holds, provided the thing you are measuring is total compensation including equity, at US technology companies.
Most employers are not that.
On base salary, the premium is closer to 10 percent
Burtch Works surveys base cash across a broad industry sample, 866 professionals in its most recent cycle. Mean base salaries for data scientists:
| Level | 2025 | 2024 |
|---|---|---|
| IC-1 | $109,545 | $99,217 |
| IC-2 | $139,837 | $130,917 |
| IC-3 | $158,861 | $157,966 |
| MG-1 | $160,869 | $149,473 |
| MG-2 | $192,507 | $181,590 |
| MG-3 | $247,000 | $227,077 |
And the finding that matters: "AI professionals still command a 9 to 13% cash premium over Data Scientists."
Apply that to a senior individual contributor. IC-3 at $158,861 plus 9 to 13% lands between roughly $173,000 and $179,500. That is a materially different conversation from $278,125.
Why the two disagree
Three reasons, and each one changes how you should use the number.
One counts equity, the other does not. levels.fyi sums base, stock and bonus. Burtch Works reports base cash. At the companies that dominate levels.fyi submissions, stock is the larger half of the package. If your business does not grant meaningful equity, the levels.fyi gap is not a gap you can close, and not one you need to.
The samples are not the same population. levels.fyi is voluntary self-report, skewed towards large US technology firms and frontier AI labs, and skewed further by who bothers to submit, which is disproportionately people pleased with their offer. Burtch Works surveys across insurance, healthcare, manufacturing and retail. One is measuring the tail. The other is measuring the body.
The titles are not stable. "AI engineer" currently covers everything from wiring a retrieval pipeline into an existing product to pre-training research. Any survey that groups those together produces a median that describes nobody.
The UK gap is smaller again, and nobody should sound confident about it
| Source | Data scientist | ML engineer | Gap |
|---|---|---|---|
| Robert Half 2026, London median | £88,750 | £102,000 | +15% |
| Robert Half 2026, UK-wide range | £56,250 to £81,000 | £60,000 to £95,000 | +7% to +17% |
| IT Jobs Watch, UK median, 6 months to 10 Aug 2026 | £75,000 | £76,000 | +1.3% |
Robert Half's London 75th percentiles are £129,250 for ML engineers against £110,250 for data scientists, so the gap widens slightly at the top of the band but never approaches 30%.
The IT Jobs Watch figures deserve a warning in both directions. That source shows the ML engineer median down 10.59% year on year while data scientist is up 7.14%. Do not read that as a collapse in ML pay. It rests on 36 quoted salaries against 376 for data scientist, which is noise, not a trend. What it does establish is that the advertised UK permanent market shows no reliable ML premium at all, and anyone quoting you a widening UK gap is extrapolating.
Third UK view, and it disagrees again: Glassdoor puts the London ML engineer average at £75,677 from 828 reports as of May 2026, below Robert Half's 25th percentile of £81,500. The reason is structural: Glassdoor blends all seniorities and company sizes including junior and non-London-weighted respondents, while Robert Half prices the roles it actually places, which skew to funded London firms. Neither is wrong. They are different questions.
For a total compensation view, levels.fyi puts the London data scientist median at £89,341, with a 90th percentile of £161,000 and JPMorgan Chase top of the table at £144,929.
The PwC 62 percent number is measuring something else entirely
You will see PwC's 2026 Global AI Jobs Barometer cited in these conversations: a 62% wage premium for AI skills, up from 56%, drawn from over a billion job adverts across 27 countries.
That is a premium for AI skills appearing in an advert, in any occupation, against a non-AI advert in the same occupation. It is not a data scientist versus ML engineer comparison. PwC's own range runs from 118% in consumer markets to 16% in government, which is a spread wide enough to tell you the average is not describing your role. It is a useful macro signal. It is not a banding input.
What you are actually paying for
Strip out the measurement noise and the durable premium attaches to production ownership, not to model-building. Burtch Works describes the shift plainly: scarcity has moved from research-flavoured work to applied LLM engineering, retrieval and data quality, evaluation and safety, and model governance. You are paying for someone who owns a system after it ships.
And the extreme numbers driving the discourse come from a cohort you probably do not compete with. H-1B disclosures showed Anthropic base pay for some technical staff reaching $1.38m and $1.12m, and that is base only, before equity. Those roles pull every US median upward. They are not your comparator, and treating them as one will cost you a great deal for no additional retention.
Setting a band this quarter
One takeaway: band on base cash first, then decide separately and explicitly whether you are buying production ownership, and price that at 10 to 15%, not 40%.
In practice, for a senior individual contributor. In London, anchor to Robert Half's ML engineer 50th to 75th percentile, £102,000 to £129,250, and expect to sit in the lower half of it unless you grant real equity. In the US, start from the Burtch Works IC-3 base of $158,861, add the evidenced 9 to 13% AI premium to reach roughly $173,000 to $180,000, and move towards levels.fyi's $278,125 only if you are genuinely competing with equity-heavy technology employers for the same candidates.
If a candidate brings you a levels.fyi screenshot, the correct response is not to match it. It is to ask what proportion of that figure was stock, and whether they realised it.
If you are rebuilding a data and AI band this quarter and want the underlying figures pressure-tested against what is actually being accepted in your market, talk to Synerjy.