This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more on smartphones and digital imagery — real or otherwise — follow Allison Johnson. The Stepback arrives in our subscribers’ inboxes at 8AM ET. Opt in for The Stepback here.
Artificial Intelligence
AI image generators are getting better by getting worse
Remember the early days of AI image generation? Oh how we laughed when our prompts resulted in people with too many fingers, rubbery limbs, and other details easily pointing to fakes. But if you haven’t been keeping up, I regret to inform you that the joke is over. AI image generators are getting way better at creating realistic fakes, partly thanks to a surprising new development: making image quality a little bit worse.
If you can believe it, OpenAI debuted its image generation tool DALL-E a little less than five years ago. In its first iteration, it could only generate 256 x 256 pixel images; tiny thumbnails, basically. A year later, DALL-E 2 debuted as a huge leap forward. Images were 1024 x 1024, and surprisingly real-looking. But there were always tells.
In Casey Newton’s hands-on with DALL-E 2 just after it launched in beta, he included an image made from his prompt: “A shiba inu dog dressed as a firefighter.” It’s not bad, and it might fool you if you saw it at a glance. But the contours of the dog’s fur are fuzzy, the patch on its (adorable little) coat is just some nonsense scribbles, and there’s a weird, chunky collar tag hanging to the side of the dog’s neck that doesn’t belong there. The cinnamon rolls with eyes from the same article were easier to believe.
Midjourney and Stable Diffusion also came to prominence around this time, embraced by AI artists and people with, uh, less savory designs. New, better models emerged over the next couple of years, minimizing the flaws and adding the ability to render text somewhat more accurately. But most AI generated images still carried a certain look: a little too smooth and perfect, with a kind of glow you’d associate with a stylized portrait more than a candid photo. Some AI images still look that way, but there’s a new trend toward actual realism that tones down the gloss.
OpenAI is a relative newcomer in the tech world when you compare it to the likes of Google and Meta, but those established companies haven’t been standing still as AI ascends. In the latter half of 2025, Google released a new image model in its Gemini app called Nano Banana. It went viral when people started using it to make realistic figurines of themselves. My colleague Robert Hart tried out the trend and noticed something interesting: the model preserved his actual likeness more faithfully than other AI tools.
That’s the thing about AI images: they often tend toward a neutral, bland middle ground. Your request for an image of a table will look basically right, but it will also feel like the result of a computer averaging out every table it’s ever seen into something lacking any actual character. The things that make an image of a table look like the real thing — or a reproduction of your own facial features — are actually imperfections. I don’t mean the bizarre artifacts of AI trying to understand letters of the alphabet. I mean a little clutter, messiness, and lighting that’s less than ideal. And lately, that also means imitating the imperfections of our most popular cameras.
Google updated its image model less than a month ago, touting Nano Banana Pro as its most advanced and realistic model yet. It’s able to draw from real-world knowledge and render text better, but the thing I find most interesting is that it often mimics the look of a photo taken with a phone camera. Contrast (or lack thereof), perspective, aggressive sharpening, exposure choices — so many of the images this model generated for me bear the hallmarks of phone camera systems.
Whether you’re aware of it or not, you’re probably attuned to this look, too. The small sensors and lenses in our phones use multiframe processing to overcome their limitations compared to a bigger camera, and these photos are optimized for viewing on a smaller screen. Altogether, that means phone photos have a certain “look” compared to a more artistic representation of a scene — boosting shadows to reveal more details and cranking up sharpness to make subjects pop. Apparently, Google’s image generator has absorbed this style, too.
Google isn’t alone in offering a more realistic look to generated images. Adobe’s Firefly image generator has a control labeled “Visual Intensity” that lets you tone down the glowy AI look. The results look less pristine and more like they were captured with a real camera — maybe more of a professional camera than a phone camera, which makes sense given Adobe’s target audience of professionals. But even Meta’s AI generator has a slider for “Stylization,” which dials the realism up or down accordingly. Elsewhere, video generation tools like OpenAI’s Sora 2 and Google’s Veo 3 have been used to create viral clips mimicking the low-resolution, grainy visuals of security cameras. When the AI only has to be as good as a CCTV, it can be pretty convincing.
There are a lot of good reasons to treat claims of AI’s infinite potential for improvement with skepticism. AI agents still struggle at buying you a pair of shoes. But the imaging models? They have vastly improved, and the evidence is in front of our eyes.
I recently spoke to Ben Sandofsky, one of the cofounders of the popular iPhone camera app Halide, about the AI-imitating-smartphones trend recently. He says that by embracing the strong processing tendencies and familiarity of phone camera photos, which already make our photos look a little untethered from reality, “Google might have sidestepped around the uncanny valley.” AI doesn’t have to make a scene look realistic — in a way, that’s a dead giveaway. It just has to mimic the way we record reality, with all its flaws, and use it as a kind of cheat code to make an image look believable. So how do we believe any photo that we see?
There’s the Sam Altman view, that real imagery and AI imagery will blend together in the future, and we’ll just be fine with that. I think he’s partially right, but I have a hard time believing that we won’t really care what’s real and what’s not. And in order to sort the two out for ourselves, we’re going to need some help. And it appears to be on the way — but it’s not coming as fast as the AI image models are improving.
The C2PA’s Content Credentials standard is gaining some much-needed momentum. On Google’s Pixel 10 series phones, every image taken with the camera gets a cryptographic signature identifying how it was made. This avoids the “implied truth effect,” as Pixel camera head Isaac Reynolds explained to me earlier this year. If you only label AI-generated images as AI, then we assume that everything without a label is real. Actually, though, the lack of a label only means that we don’t know where the image came from. So the Pixel camera labels both AI and non-AI images alike.
Labels are all well and good, but they’re not useful if you can’t see them. That’s starting to change, and earlier this year Google Photos added support for displaying Content Credentials. The company will also make Content Credentials easy to view in search results and ads when they’re present. That last part is the key, though — right now, most images captured with phone cameras today aren’t assigned credentials. For the system to work, hardware makers need to adopt the standard so that images are marked as AI or not at the point when they’re created. The platforms where images are shared need to get on board, too. Until that happens, we’re on our own — and it’s a better time than ever to trust nothing that you see.
- Google’s Pixel 10 cameras don’t just offer AI image editing tools — there’s a generative AI model baked right into the imaging pipeline. It’s only used in a feature called Pro Res Zoom, and it aims to improve on what would otherwise be pretty crappy digital zoom image quality. It doesn’t work on people for now, which is a good thing in my book.
- Traditional camera makers are adopting C2PA’s Content Credentials as well, albeit slowly, like the $9,000+ Leica M-11P.
- Meanwhile, AI-powered editing tools in Photoshop like generative fill have become more powerful and popular with photographers. There’s a middle ground between fully AI-generated images and photos untouched by AI that’s getting trickier to define.
- My colleague Jess Weatherbed wrote a great explainer of C2PA that’s (frustratingly!) still a good reflection of where we are a year later.
- Wired talked to Google’s Pixel camera team around the Pixel 9 launch about how it treats our photos like memories.
- Bloomberg investigated the community of creators using tools like Sora 2 to create AI generated slop for kids on YouTube. Bleak!
Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.
Continue Reading
Artificial Intelligence
Ronnie Sheth, CEO, SENEN Group: Why now is the time for enterprise AI to ‘get practical'
Before you set sail on your AI journey, always check the state of your data – because if there is one thing likely to sink your ship, it is data quality.
Gartner estimates that poor data quality costs organisations an average of $12.9 million each year in wasted resources and lost opportunities. That’s the bad news. The good news is that organisations are increasingly understanding the importance of their data quality – and less likely to fall into this trap.
That’s the view of Ronnie Sheth, CEO of AI strategy, execution and governance firm SENEN Group. The company focuses on data and AI advisory, operationalisation and literacy, and Sheth notes she has been in the data and AI space ‘ever since [she] was a corporate baby’, so there is plenty of real-world experience behind the viewpoint. There is also plenty of success; Sheth notes that her company has a 99.99% client repeat rate.
“If I were to be very practical, the one thing I’ve noticed is companies jump into adopting AI before they’re ready,” says Sheth. Companies, she notes, will have an executive direction insisting they adopt AI, but without a blueprint or roadmap to accompany it. The result may be impressive user numbers, but with no measurable outcome to back anything up.
Even as recently as 2024, Sheth saw many organisations struggling because their data was ‘nowhere where it needed to be.’ “Not even close,” she adds. Now, the conversation has turned more practical and strategic. Companies are realising this, and coming to SENEN Group initially to get help with their data, rather than wanting to adopt AI immediately.
“When companies like that come to us, the first course of order is really fixing their data,” says Sheth. “The next course of order is getting to their AI model. They are building a strong foundation for any AI initiative that comes after that.
“Once they fix their data, they can build as many AI models as they want, and they can have as many AI solutions as they want, and they will get accurate outputs because now they have a strong foundation,” Sheth adds.
With breadth and depth in expertise, SENEN Group allows organisations to right their course. Sheth notes the example of one customer who came to them wanting a data governance initiative. Ultimately, it was the data strategy which was needed – the why and how, the outcomes of what they were trying to do with their data – before adding in governance and providing a roadmap for an operating model. “They’ve moved from raw data to descriptive analytics, moving into predictive analytics, and now we’re actually setting up an AI strategy for them,” says Sheth.
It is this attitude and requirement for practical initiatives which will be the cornerstone of Sheth’s discussion at AI & Big Data Expo Global in London this week. “Now would be the time to get practical with AI, especially enterprise AI adoption, and not think about ‘look, we’re going to innovate, we’re going to do pilots, we’re going to experiment,’” says Sheth. “Now is not the time to do that. Now is the time to get practical, to get AI to value. This is the year to do that in the enterprise.”
Watch the full video conversation with Ronnie Sheth below:
Greg Holmes, Field CTO for EMEA at Apptio, an IBM company, argues that successfully scaling intelligent automation requires financial rigour.
The “build it and they will come” model of technology adoption often leaves a hole in the budget when applied to automation. Executives frequently find that successful pilot programmes do not translate into sustainable enterprise-wide deployments because initial financial modelling ignored the realities of production scaling.
“When we integrate FinOps capabilities with automation, we’re looking at a change from being very reactive on cost management to being very proactive around value engineering,” says Holmes.
This shifts the assessment criteria for technical leaders. Rather than waiting “months or years to assess whether things are getting value,” engineering teams can track resource consumption – such as cost per transaction or API call – “straight from the beginning.”
The unit economics of scaling intelligent automation
Innovation projects face a high mortality rate. Holmes notes that around 80 percent of new innovation projects fail, often because financial opacity during the pilot phase masks future liabilities.
“If a pilot demonstrates that automating a process saves, say, 100 hours a month, leadership thinks that’s really successful,” says Holmes. “But what it fails to track is that the pilot sometimes is running on over-provisioned infrastructure, so it looks like it performs really well. But you wouldn’t over-provision to that degree during a real production rollout.”
Moving that workload to production changes the calculus. The requirements for compute, storage, and data transfer increase. “API calls can multiply, exceptions and edge cases appear at volume that might have been out of scope for the pilot phase, and then support overheads just grow as well,” he adds.
To prevent this, organisations must track the marginal cost at scale. This involves monitoring unit economics, such as the cost per customer served or cost per transaction. If the cost per customer increases as the customer base grows, the business model is flawed.
Conversely, effective scaling should see these unit costs decrease. Holmes cites a case study from Liberty Mutual where the insurer was able to find around $2.5 million of savings by bringing in consumption metrics and “not just looking at labour hours that they were saving.”
However, financial accountability cannot sit solely with the finance department. Holmes advocates for putting governance “back in the hands of the developers into their development tools and workloads.”
Integration with infrastructure-as-code tools like HashiCorp Terraform and GitHub allows organisations to enforce policies during deployment. Teams can spin up resources programmatically with immediate cost estimates.
“Rather than deploying things and then fixing them up, which gets into the whole whack-a-mole kind of problem,” Holmes explains, companies can verify they are “deploying the right things at the right time.”
When scaling intelligent automation, tension often simmers between the CFO, who focuses on return on investment, and the Head of Automation, who tracks operational metrics like hours saved.
“This translation challenge is precisely what TBM (Technology Business Management) and Apptio are designed to solve,” says Holmes. “It’s having a common language between technology and finance and with the business.”
The TBM taxonomy provides a standardised framework to reconcile these views. It maps technical resources (such as compute, storage, and labour) into IT towers and further up to business capabilities. This structure translates technical inputs into business outputs.
“I don’t necessarily know what goes into all the IT layers underneath it,” Holmes says, describing the business user’s perspective. “But because we’ve got this taxonomy, I can get a detailed bill that tells me about my service consumption and precisely which costs are driving it to be more expensive as I consume more.”
Addressing legacy debt and budgeting for the long-term
Organisations burdened by legacy ERP systems face a binary choice: automation as a patch, or as a bridge to modernisation. Holmes warns that if a company is “just trying to mask inefficient processes and not redesign them,” they are merely “building up more technical debt.”
A total cost of ownership (TCO) approach helps determine the correct strategy. The Commonwealth Bank of Australia utilised a TCO model across 2,000 different applications – of various maturity stages – to assess their full lifecycle costs. This analysis included hidden costs such as infrastructure, labour, and the engineering time required to keep automation running.
“Just because of something’s legacy doesn’t mean you have to retire it,” says Holmes. “Some of those legacy systems are worth maintaining just because the value is so good.”
In other cases, calculating the cost of the automation wrappers required to keep an old system functional reveals a different reality. “Sometimes when you add up the TCO approach, and you’re including all these automation layers around it, you suddenly realise, the real cost of keeping that old system alive is not just the old system, it’s those extra layers,” Holmes argues.
Avoiding sticker shock requires a budgeting strategy that balances variable costs with long-term commitments. While variable costs (OPEX) offer flexibility, they can fluctuate wildly based on demand and engineering efficiency.
Holmes advises that longer-term visibility enables better investment decisions. Committing to specific technologies or platforms over a multi-year horizon allows organisations to negotiate economies of scale and standardise architecture.
“Because you’ve made those longer term commitments and you’ve standardised on different platforms and things like that, it makes it easier to build the right thing out for the long term,” Holmes says.
Combining tight management of variable costs with strategic commitments supports enterprises in scaling intelligent automation without the volatility that often derails transformation.
IBM is a key sponsor of this year’s Intelligent Automation Conference Global in London on 4-5 February 2026. Greg Holmes and other experts will be sharing their insights during the event. Be sure to check out the day one panel session, Scaling Intelligent Automation Successfully: Frameworks, Risks, and Real-World Lessons, to hear more from Holmes and swing by IBM’s booth at stand #362.
See also: Klarna backs Google UCP to power AI agent payments
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
FedEx is using AI to change how package tracking and returns work for large enterprise shippers. For companies moving high volumes of goods, tracking no longer ends when a package leaves the warehouse. Customers expect real-time updates, flexible delivery options, and returns that do not turn into support tickets or delays.
That pressure is pushing logistics firms to rethink how tracking and returns operate at scale, especially across complex supply chains.
This is where artificial intelligence is starting to move from pilot projects into daily operations.
FedEx plans to roll out AI-powered tracking and returns tools designed for enterprise shippers, according to a report by PYMNTS. The tools are aimed at automating routine customer service tasks, improving visibility into shipments, and reducing friction when packages need to be rerouted or sent back.
Rather than focusing on consumer-facing chatbots, the effort centres on operational workflows that sit behind the scenes. These are the systems enterprise customers rely on to manage exceptions, returns, and delivery changes without manual intervention.
How FedEx is applying AI to package tracking
Traditional tracking systems tell customers where a package is and when it might arrive. AI-powered tracking takes a step further by utilising historical delivery data, traffic patterns, weather conditions, and network constraints to flag potential delays before they happen.
According to the PYMNTS report, FedEx’s AI tools are designed to help enterprise shippers anticipate issues earlier in the delivery process. Instead of reacting to missed delivery windows, shippers may be able to reroute packages or notify customers ahead of time.
For businesses that ship thousands of parcels per day, that shift matters. Small improvements in prediction accuracy can reduce support calls, lower refund rates, and improve customer trust, particularly in retail, healthcare, and manufacturing supply chains.
This approach also reflects a broader trend in enterprise software, in which AI is being embedded into existing systems rather than introduced as standalone tools. The goal is not to replace logistics teams, but to minimise the number of manual decisions they need to make.
Returns as an operational problem, not a customer issue
Returns are one of the most expensive parts of logistics. For enterprise shippers, particularly those in e-commerce, returns affect warehouse capacity, inventory planning, and transportation costs.
According to PYMNTS, FedEx’s AI-enabled returns tools aim to automate parts of the returns process, including label generation, routing decisions, and status updates. Companies that use AI to determine the most efficient return path may be able to reduce delays and avoid returning things to the wrong facility.
This is less about convenience and more about operational discipline. Returns that sit idle or move through the wrong channel create cost and uncertainty across the supply chain. AI systems trained on past return patterns can help standardise decisions that were previously handled case by case.
For enterprise customers, this type of automation supports scale. As return volumes fluctuate, especially during peak seasons, systems that adjust automatically reduce the need for temporary staffing or manual overrides.
What FedEx’s AI tracking approach says about enterprise adoption
What stands out in FedEx’s approach is how narrowly focused the AI use case is. There are no broad claims about transformation or reinvention. The emphasis is on reducing friction in processes that already exist.
This mirrors how other large organisations are adopting AI internally. In a separate context, Microsoft described a similar pattern in its article. The company outlined how AI tools were rolled out gradually, with clear limits, governance rules, and feedback loops.
While Microsoft’s case focused on knowledge work and FedEx’s on logistics operations, the underlying lesson is the same. AI adoption tends to work best when applied to specific activities with measurable results rather than broad promises of efficiency.
For logistics firms, those advantages include fewer delivery exceptions, lower return handling costs, and better coordination between shipping partners and enterprise clients.
What this signals for enterprise customers
For end-user companies, FedEx’s move signals that logistics providers are investing in AI as a way to support more complex shipping demands. As supply chains become more distributed, visibility and predictability become harder to maintain without automation.
AI-driven tracking and returns could also change how businesses measure logistics performance. Companies may focus less on delivery speed and more on how quickly issues are recognised and resolved.
That shift could influence procurement decisions, contract structures, and service-level agreements. Enterprise customers may start asking not just where a shipment is, but how well a provider anticipates problems.
FedEx’s plans reflect a quieter phase of enterprise AI adoption. The focus is less on experimentation and more on integration. These systems are not designed to draw attention but to reduce noise in operations that customers only notice when something goes wrong.
(Photo by Liam Kevan)
See also: PepsiCo is using AI to rethink how factories are designed and updated
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
-
Fintech6 months agoRace to Instant Onboarding Accelerates as FDIC OKs Pre‑filled Forms | PYMNTS.com
-
Cyber Security7 months agoHackers Use GitHub Repositories to Host Amadey Malware and Data Stealers, Bypassing Filters
-
Fintech6 months ago
DAT to Acquire Convoy Platform to Expand Freight-Matching Network’s Capabilities | PYMNTS.com
-
Fintech5 months agoID.me Raises $340 Million to Expand Digital Identity Solutions | PYMNTS.com
-
Artificial Intelligence7 months agoNothing Phone 3 review: flagship-ish
-
Artificial Intelligence7 months agoThe best Android phones
-
Fintech4 months agoTracking the Convergence of Payments and Digital Identity | PYMNTS.com
-
Fintech7 months agoIntuit Adds Agentic AI to Its Enterprise Suite | PYMNTS.com