Photoroom AI image test highlights value of freelance judgement

Photoroom tested 4,250 AI-generated product images and found that even the strongest model passed its full product-fidelity assessment only 29% of the time, highlighting the quality-control work that remains for freelance creatives.

The findings feature in advice published by Startups.co.uk on how freelancers can develop skills as AI takes on more production tasks. Contributors recommend using existing client demand to guide training, rather than giving up paid work or buying qualifications before establishing whether customers need them.

What happened

Matt Rouif, chief executive and co-founder of Photoroom, points to the test as evidence that generating an image and delivering a usable commercial asset are different tasks. For the strongest model, 71% of images did not meet the full product-fidelity standard.

His argument is that clients can now produce large numbers of images themselves, but still need someone capable of selecting suitable results. Identifying a failure, and understanding why an image falls short, creates work beyond the initial generation process.

Careers expert Drew Povey advises freelance creatives against trying to beat AI on output or turnaround time. He instead identifies judgement, creative ability, taste, client relationships and specialist knowledge as the capabilities around which to build a service.

That places responsibility for the finished result with the freelancer, even where a tool has produced the initial material. The distinction is between selling production alone and taking responsibility for whether the work meets a client’s requirements.

Software engineer Quentin Claudet, who has developed AI skills, says the technology has not removed his role in thinking through work. He nevertheless cautions against becoming too absorbed in the tool or relying on it excessively.

The background

AI’s ability to produce text, images and draft code quickly has raised concerns among freelancers about competition from cheaper, faster alternatives. The reskilling question therefore covers both which services to develop and how to fund the learning period without interrupting earnings.

Povey proposes assessing prospective skills against three factors: existing strengths, customers’ willingness to pay and personal interest. His view is that a direction supported by all three is more durable than one chosen because a tool or job title is attracting attention.

Accountant Dat Ngo recommends examining invoices before consulting trend reports. Repeat customers, prompt payment and limited disputes can help identify the relationships worth developing, rather than treating every part of a freelance client base as equally attractive.

Moving into a completely unrelated service carries two demands at once: learning the work and finding a new market for it. Ngo favours related capabilities that can be sold to customers a freelancer already serves, avoiding the need to rebuild both expertise and client acquisition simultaneously.

Connor Gillivan suggests looking for work that clients repeatedly request but cannot get done. Harvey Dhillon, founder and chief executive of a UK accountancy firm, similarly regards services already generating revenue as a relatively low-cost starting point for deciding what to learn.

For those without an extensive invoice history, accountant Andrew Gosselin recommends direct conversations. He distinguishes the clients producing the best effective hourly return from those placing the largest orders, then suggests asking what additional capability would make the freelancer more useful to those customers.

The focus on practical applications also appears in Emma Jones’s call for small firms to trial AI on late payments, which puts an existing business problem at the centre of technology adoption.

What people are saying

Ngo advises freelancers to retain 70% to 80% of their paid workload during training. Before cutting that workload, he recommends holding cash equal to at least three months of essential expenses, putting a financial buffer ahead of a substantial change in working pattern.

Gosselin argues against lowering prices while learning. In his view, a reduced rate can imply that earlier work was overpriced; reducing the amount of work accepted is preferable to changing the price attached to established expertise.

Povey also warns against assuming that an expensive course or qualification will produce a new career. His recommendation is to establish demand through client conversations before committing money, rather than treating the certificate itself as evidence of a viable service.

Rouif draws a distinction between learning a particular platform and developing skills that survive changes to it. He considers narrowly tool-specific knowledge vulnerable to the next model release, while accountability for the finished output remains relevant across platforms.

Povey’s further concern is that freelancers may discard assets they have already built. Reputation, professional contacts and industry understanding should support the transition, he argues, rather than being set aside in an attempt to make a complete break with existing work.

What happens next

Rouif recommends beginning with a recurring assignment where the freelancer already understands the quality standard. Using a task performed each week creates an opportunity to practise a new capability within paid work, with an established basis for judging whether the result is good enough.

Gillivan proposes a more structured trial with one existing customer: add a related service, such as strategy, alongside the normal assignment for a quarter. That approach tests the proposed offer on a live account rather than postponing commercial experience until training has finished.

He says the trial can produce a case study and may lead to a higher rate with the same client. Those are potential outcomes, not guaranteed returns; the immediate exercise is to establish whether the additional service meets a real customer need.