Services

About

Work

Pricing

Book

Charcoal art with messy hands on the canvas

Where I land on AI

I started Melty with the aim of making authentic storytelling and design more available to local businesses doing great work. This followed a six-year stint in corporate tech that had admittedly left me both burnt out and disillusioned.

Part of that disillusionment was watching where the tooling was headed. The clear direction of travel, across the whole category, was toward generating the content for you: fewer decisions, less involvement, more volume. Social media as a box to check rather than a space to own authentically. There's certainly a market for that. Perhaps folks who realize a digital presence is important to their business and want to handle it with as little effort as possible. It didn't resonate with me, and by the end I didn't want to spend my working life building it.

Social media apps aren't exactly paragons of virtue these days, so I get the desire to game the system a little, but I think this type of content reflects poorly on both businesses and the agencies claiming to support them.

Most consumers' first encounter with a business is online, which has made how you show up there more important than ever. Businesses that already provide a compelling offering and experience in person have a great opportunity to align the digital side of their business with that momentum, but it's a lot of work to do well. Website, social media, listing platforms, brand design crossing physical and digital mediums, etc, etc.

So the landscape as I see it: a big opportunity to excel, big rewards and differentiation for businesses that do, and a fair amount of work to get there.

Lucky for me, I really like this kind of work and I love the businesses I get to work with.

I've already wandered off track though. This article is about AI, and hopefully providing those who are interested in teaming up with me, or in using AI for their own businesses, a deeper window into my approach and perspective.


Cries from both sides

Part of why I'm writing this is because it's become inconvenient for me to remain ambivalent on the topic.

I swim in a few different circles, which is great, I like hearing different perspectives on the world. A little more troubling though: I've noticed many voices trending toward one of two polar extremes.

From folks in the tech space: "Dude you can do SO much with AI. Are you using it to optimize your life? Are you using it to be more efficient? You should ask AI for a review of your relationships! AI is awesome! AI is going to replace you! Be excited! Be scared! Don't forget to invest in AI (just like how I told you to invest in crypto). etc, etc"

From social justice oriented folks: "AI is evil. AI uses water, did you know that? The data centres make noise? Did you know that? I think it's so stupid. AI could never be useful. AI is everything that's wrong with the world right now. etc, etc"

Of course not everyone is at these extremes, but I get questions about it from all perspectives (and from comments on clients' posts: "is this ai?") so I figured I'd better take the time to jot down my own thoughts.


A brief history of AI

AI has been around for a long time, and it's been hyped and abandoned twice before this.

The first wave ran from the 1950s through the early '70s, when the field was confident that human-level machine reasoning was a decade or two out. When the funding agencies noticed it wasn't, the money left. The second wave came in the 1980s around "expert systems," software that encoded a specialist's decision-making into a big pile of rules. It worked well enough in narrow domains that corporations bought in hard, then discovered the systems were expensive, brittle, and nearly impossible to keep current. That collapse kicked off what's usually called the second AI winter, and it lasted the better part of two decades.

The technology didn't disappear during those winters, it just stopped being called AI. It got embedded in things and renamed. In 1990, Matsushita launched a fuzzy-logic washing machine in Japan called "Beloved Wife Day Fuzzy," and sold it explicitly on the promise of machine intelligence: hundreds of wash cycles, one button, the machine figures out the rest. Fuzzy logic ended up in rice cookers, camcorders, vacuum cleaners and subway braking systems. A lot of cities now run adaptive traffic signals that adjust timing from live sensor data rather than a fixed cycle. Your car is full of adaptive control systems, and increasingly of actual machine learning if it has lane-keeping or collision avoidance.

None of that is what people mean when they say "AI" today. When people levy criticism at the technology they're typically referring to this iteration of chatbots and generative tools that came into vogue circa 2022/23. That's a fair distinction to make and I'll stick to it for the rest of this piece, but it's worth knowing that we've been here before, twice, and that the pattern has been: enormous promises, real but narrower capability, a correction, and then the useful parts quietly get absorbed into everything.

If you want a good long-form look at this history, The Last Invention from Longview is worth checking into. It's real journalism rather than straight hype, and it takes the doomers, the skeptics, and the accelerationists each seriously in turn.


The two voices of "AI bad!"

There are two versions of the critique I hear most, and they don't have much to do with each other.

1. AI is killing the planet.

The specific claims that circulate here are usually wrong, and the underlying concern is usually right, which is a frustrating combination to argue about.

The "one bottle of water per query" line you've seen on Instagram traces back to a 2023 University of California, Riverside study, and it's been mangled in transmission. The number described water spread across a long response and a full session, not a sip taken per question. Google published a detailed methodology in 2025 putting a median Gemini text prompt at 0.26 millilitres, roughly five drops, alongside about as much energy as nine seconds of television. Independent estimates land anywhere from a fraction of a millilitre to a few tens of millilitres depending on the model and on what you count. So the per-prompt panic is misplaced.

The aggregate picture is a different story. Researchers project that US data centres could be consuming somewhere between 731 billion and 1.1 trillion litres of water annually by 2030; the top of that range is roughly New York City's yearly drinking water supply. Texas alone is projected to go from tens of billions of gallons a year to potentially hundreds of billions by 2030. And all water is local, it doesn't help a county in Georgia or Oregon that the global average looks fine.

The noise complaint is real too, and I'd stop dismissing it if I were on that side of the argument. There's been a string of class actions filed in 2026 in Michigan, New Jersey, Mississippi and Wisconsin over the constant low-frequency hum of cooling systems and gas turbines, with residents measuring 60 to 70 decibels at the property line. That's a dishwasher running twenty-four hours a day, outside your house, forever. That's somebody's actual life.

Where I come out: the energy, water, and climate intensity of this stuff is a real problem, it is very unevenly distributed, and it is at least partly an engineering and siting problem rather than an inherent property of the technology. There are meaningfully better and worse ways to build this infrastructure. China, for example, now requires new large data centres to hit a power usage effectiveness below 1.25, which is aggressive, and it's adding renewable capacity at a scale nobody else is matching. It's also true that Chinese data centres still draw something like 60 to 70 percent of their power from coal, so I'm not holding anyone up as the green example here. The honest version is that efficiency mandates work, siting decisions matter enormously, and almost nobody is being held to either standard yet.

2. AI is going to kill us all.

I find this one harder to be flippant about than I'd like to be, largely because the people making the argument aren't quacks. Geoffrey Hinton and Yoshua Bengio, two of the three people most responsible for the deep learning techniques underneath all of this, have both publicly said they're worried. That deserves more than an eye roll. Bill Gates has reportedly switched his tone recently too, but I haven’t looked into that much yet..

That said, I don't think what we have right now is a mind. What's in front of us today is an extraordinarily good pattern machine, and the distance between that and something with intentions of its own looks large to me. Where I do get nervous is less about the machine waking up and more about people handing it authority it hasn't earned, and building systems that fail confidently instead of loudly.

Either way, it's a fair distance from the question a local business owner or their customers is actually asking me, which is whether their Instagram content was produced by a data-centre.


Big promises, big hype

I bring this up every time the topic of AI is broached because I think it’s very important to understand motivations.

The current AI industry isn't just funded, it's funded on a scale that requires it to be world-historically transformative in order to break even. Bain & Company's 2025 global technology report put a number on it: meeting projected AI demand implies roughly $500 billion a year in data centre capital expenditure by 2030, which in turn requires about $2 trillion in annual AI revenue to be profitable. Their own generous-case math still leaves the industry about $800 billion a year short.

To be worth what investors have already paid for it, this technology has to be valuable enough to cure cancer, solve climate change, and end world hunger. That is roughly the size of the hole. And it explains a lot of the noise, because everyone with money in the game needs you to believe the largest possible version of the story.

Meanwhile, on the ground: MIT's NANDA initiative surveyed enterprise adoption in 2025 and found that around 95 percent of generative AI pilots produced no measurable return. The report is preliminary and has been fairly criticized for a narrow definition of success, but the direction of it matches what I saw from the inside. The failures weren't about model quality. They were about companies buying a tool before they understood the problem, and then quietly shelving it.

I don't say this to dunk on the technology. I say it because it means you should be skeptical of anyone, including me, telling you AI will transform your business. Most of the money currently being spent on that promise is not producing anything.


Threading the needle: where I land

The primary course load of my interactive systems design degree was coding. Coding is cool, but the way they taught it was a bit dated. I wrote most of my exams with pencil and paper, no IDE, memorizing proper syntax.

It's probably no surprise to my peers that I gravitated toward design precisely because I found the rationale, the psychology, and the art of it all more compelling than the ones and zeros behind the scenes.

All that to say, I use AI for coding a lot. It's raised the ceiling on what's possible for one person, and especially for designers. Probably cringe for some of my software peers to hear this, but I've appreciated being able to do more without spending hours in Stack Overflow.

Where I draw the line is on storytelling.

My entire approach is rooted in truthful and compelling storytelling across digital and physical spaces, so I'm critical when it comes to using AI to tell stories, whether photographic, video, or written. If I'm asking someone to take the time to engage with real work from another human being, that avenue deserves that I take the time to pave it myself, honestly. A generated photo of a dish you kinda serve, in a room that doesn't exist, staffed by people who don't work there, isn't a shortcut to the story. It's a substitute for having one.

There's also a plain business case, separate from the principle. The thing that differentiates a good local business is that it is specific and real. Generated content is, definitionally, an average of everything that came before it. Using it to represent something distinctive is working directly against the only advantage you have.


What people actually think about this

I'd been making that argument on instinct for a while before I went looking to see whether the data backed it up. It does, fairly convincingly.

The blunt version comes from a December 2025 study by Klaviyo and Datalily: 7 percent of consumers said visible AI-generated marketing made them trust a brand more, and 31 percent said it made them trust the brand less. That's better than four to one against.

There's also proper experimental work on this, not just polling. Researchers at the Nuremberg Institute for Market Decisions ran controlled experiments across the US, UK and Germany where they showed people identical ads, telling one group the image was a photograph and the other group it was AI-generated. Same ad, only the label changed. The AI-labelled version was rated as less appealing, less credible, and notably weaker on emotional impact; and people were measurably less likely to click through. A peer-reviewed study in the Journal of Retailing and Consumer Services found the same effect specifically for social media content, and identified the mechanism: it runs through perceived brand authenticity. People don't consciously object to the tool. They just quietly stop believing you. People don’t like inauthenticity.

And the direction of travel isn't toward acceptance. The agency Billion Dollar Boy found consumer preference for AI-generated creator content dropped from 60 percent in 2023 to 26 percent by early 2026. Whatever novelty was doing early on, it's spent.

Now, the honest complication, because I promised myself I'd include the parts that cut against me. The Nuremberg researchers titled their write-up "Transparency Without Trust" for a reason: the disclosure I described above is not free. Labelling content as AI-generated is precisely what triggers the penalty. Their conclusion was that transparency "reveals a fundamental problem but doesn't solve it." So my policy of saying so out loud costs something real, and I'd be lying if I framed it as a clever growth tactic. It's just the version I can live with. The alternative is hoping nobody notices, and getting caught later is worse than being upfront now.

The second complication is stranger. In that same research, only about a quarter of people believed they could reliably spot AI-generated content, and the honest answer is that most of them can't. Which means the accusation flies in both directions. Real photographs get called AI. Genuinely hand-written captions get dismissed as ChatGPT because someone spotted an em-dash.

I know this one first-hand. I shot an image recently that was very highly orchestrated on the practical level, and the comments came in that it looked like AI. Too golden. Too perfect. What turned them around wasn't a denial, it was the next frame in the carousel: the behind-the-scenes shot, mess and all.

Saying "this is real" is a claim. Showing the work is evidence. Only one of those travels.


The jagged edge

Generated content is smooth. It has to be, it produces the most probable next thing, over and over, which lands it on the average of everything. Competent, frictionless, and nothing in it that a person had to wrestle into place.

Real work has a jagged edge. The plate is off-centre because that's where the light was. Somebody laughs at the wrong moment and you keep it because it's the best thing in the clip. Those aren't flaws on the way to polish. They're the evidence that a person with judgment was standing in the room making calls.

Consumer psychology has a name for why that lands: the handmade effect, where people assign more value to work they perceive as made by human hands, even when objective quality is held equal. Related work on the labour illusion finds that making effort visible increases how much people value the result. So the behind-the-scenes frame isn't a bonus deliverable. It's the mechanism.

Which cuts against the instinct most businesses have, to show only the most polished version. Polish is cheap now, and infinitely available. The tripod in the shot is not.


The part I can't prove

I'll flag clearly that I don't have a study for this one.

I don't think the value is in any single honest photo. I think it's in what happens when a customer senses that the same thing is true all the way down: that the care in the marketing is the care in the room, on the plate, in how they're treated at the door, and between the business walls. When authenticity is a through-line rather than a coat of paint, people feel it, even if they'd never put it in those words. And nothing ever has to be walked back later.

That's a conviction I hold and it's one I'll bet on hard, because everything about where this is going points the same way. As the feed fills up with content that's optimised and polished and pulled toward the mean, the mean gets cheaper by the day. What's left over; the specific, the particular, the thing that could only have come from you, is the only part that can't be churned out of an automation. Your competitors can buy every tool you can. They can't buy what actually happens in your business.

So when I show up the deliverable isn't only the photographs, brand assets, and a websites. It's that you end up with something nobody can fabricate: proof, with a date on it, that this happened, and that it was worth showing. It's a slower way to work and I won't pretend otherwise. I just think it's the version that keeps being worth something.

When it comes to operational tasks, where largely it's just one machine talking to another (the internet and my codebase, for example), I'm considerably more generous.

So, concretely, here's what that looks like when we work together:

Where I use it: writing and debugging code for the sites I build, technical research, wrangling structured data, transcription, first-pass drafts of things I'm going to rewrite anyway, and the tedious middle of jobs where nobody would call the output "the work."

Where I don't: photography, video, and the words that go out under your name. Your images are photographs I took. Your captions are sentences somebody wrote. If a client wants generated imagery, I'll say so out loud, on the post, rather than letting it pass as real.

That's the whole policy. It isn't a moral position I'm asking you to adopt, it's just how I work, and I'd rather you know it up front.


Thanks for reading

<3


Sources & further reading

On the history

On water and energy

On noise

On China

On the money

On what consumers think of AI-generated marketing

Small Business Branding in Saskatoon: Why Cobbling It Together Costs More ›

Menu