AI Slop Is Costing You Trust And Power. Here Is What AI Should Be Doing Instead

Generative image and video is the most expensive, least trusted thing you can point AI at. The boring back office is where the return sits.

Aidan Lambert Published 1 Sep 2026 7 min read

What AI slop actually means

AI slop is generated content published in volume because it costs almost nothing to make: images, video and text, mostly unchecked and mostly filler. The phrase spread through 2024 as the tools went mass market, and it stuck because it named something people had already started noticing in their feeds.

That is the whole definition.

The more useful question is what it costs. Not to make, since that is close to zero, which is the entire problem. What it costs in trust, and what it costs in power. Both are measurable. Both are worse than the people selling generated content tend to mention, and neither of them shows up on the invoice you approve at the end of the month.

Problem one. Nobody trusts a picture any more

The consumer numbers are blunt. The Harris Poll, with the 4As and Infillion, reported at Cannes in June 2026: 73% of people say they are less likely to trust an ad they suspect was made by AI. 63% say they are less likely to buy from the brand behind it. 78% say AI makes advertising feel less authentic, which is the number that should worry anyone whose brand rests on being believed.

Three numbers, all pointing the same way.

A larger study says the same thing from the other direction. Klaviyo, with Datalily, surveyed more than 8,000 consumers across eight markets including Australia in December 2025. Visible AI-generated marketing lifted brand trust for 7% of them. It reduced trust for 31%.

Read those two numbers next to each other. The upside is a rounding error, and the downside is more than four times larger.

That is the trade being made every time a generated hero image goes up to save a photography budget, and almost nobody running the campaign has seen it stated that way.

Cheap to make. Expensive to be anywhere near.

The Australian fraud problem this feeds

Cheap synthetic media has a victim, and here the bill is already public. ASIC removed 19,400 scams in the 2026 financial year, a 182% rise year on year, and 33,400 across the three years since it started counting.

The wider bill is bigger. The National Anti-Scam Centre put reported scam losses at $2.18 billion in 2025, with $837.7 million of that going to investment scams, the single largest category.

The delivery mechanism is a deepfaked public figure. Anthony Albanese and Gina Rinehart both appear on ASIC's list of the most impersonated Australians, alongside Alan Kohler, Jacqui Lambie and Dick Smith. ASIC chair Sarah Court put it plainly: AI is making investment scams more convincing and harder to detect, and a simple online search is not enough to verify whether an opportunity is legitimate.

That is the environment a generated brand image now lands in. Every synthetic face in an ad teaches the audience to hunt for the seam, and an audience that has been trained by a year of investment scams gets better at finding it every month.

Problem two. Images are the most expensive thing you can ask AI to do

Luccioni, Jernite and Strubell measured this properly, at FAccT in 2024. Generating an image averages about 2.9 kWh per 1,000 inferences. Classifying text averages about 0.002 kWh across the same volume. That is a spread of roughly 1,450 times. The least efficient image model in the study drew about the same power per 1,000 images as 522 smartphone charges, which is a useful way to picture a cost that otherwise arrives as somebody else's electricity bill.

That is not a rounding difference. It is a different category of thing.

The trend holds when you zoom out. The International Energy Agency puts data centre consumption at about 415 TWh in 2024, roughly 1.5% of global electricity, and projects about 945 TWh by 2030, with AI-accelerated servers responsible for almost half of that increase.

Be careful with what that does and does not say. Data centres do not exist mainly to make pictures, and anyone arguing they do is one fact-check from losing the room. The claim that survives scrutiny is narrower and more useful: per task, generative media is the most power-hungry thing on the menu, and text and structured data sit at the cheap end by three orders of magnitude.

Where the water goes

Cooling is the other cost. Data centres use water to shed heat, and the build-out around Sydney and Melbourne has turned that into a live planning argument rather than an abstract one.

Worth being honest about the numbers. Most published water figures for AI workloads are estimates derived from power draw and regional cooling assumptions, not disclosed operator data. Operators publish very little, and what they do publish is rarely broken down by workload, so the honest position is that the direction is clear and the precision is not. We would rather say that than quote a figure to two decimal places that nobody can trace back to a source.

The comparison nobody in your feed is making

Energy per 1,000 inferences by task type. Measured by Luccioni, Jernite and Strubell, FAccT 2024. The second panel is the same data on its own scale, because on the first one eight of the ten bars are thinner than a hairline.

Generated media against automated admin, per task
	Generated image or video	Automated admin task
Energy per task	Highest measured category, about 2.9 kWh per 1,000 generations	Lowest measured category, about 0.002 kWh per 1,000 text classifications
Who sees a mistake	Every customer, permanently	One reviewer, before it leaves the building
Effect on trust	Falls for 31%, rises for 7%	Invisible to the customer
Time returned	Minutes, on a task a person enjoyed	Hours, on a task nobody wanted

The argument in one line: the same compute spent on a hero image nobody trusts could clear a week of invoice matching that everybody benefits from.

Point AI at the admin instead

This is the part that does not trend, which is most of why it is still available. Six jobs, all of them dull, all of them worth more than a generated campaign.

None of it will win an award.

Inbox triage and routing. Every enquiry read, categorised and sent to the right person, with the odd one flagged for a human to decide. This is ordinary workflow automation with a model doing the reading. On a busy shared inbox that is usually about an hour a day back, and it is an hour that used to go before anyone had started on real work.

Quoting and proposal assembly. A template, the CRM record and the pricing rules, assembled into a draft in under a minute. The estimator checks it and sends it, with the record living in the CRM rather than a folder. Quoting stops being the thing that follows people home, which is most of why it gets automated first.

Invoice and receipt matching. Line items matched against purchase orders, the exceptions surfaced, the rest cleared without anyone opening a spreadsheet. Days a month in most finance functions, and the days come back to the person who was least glad to be spending them.

Meeting notes into actioned tasks. The recording becomes a list of who owes what by when, filed in the system where the work actually lives rather than a document nobody reopens.

First drafts for repeat enquiries. The same forty questions that arrive every week, answered from your own documentation rather than invented, with a person reading every one before it sends. The same pattern runs the marketing side, where the follow-up nobody sends is usually the expensive gap.

Data entry between systems that were never meant to talk. The copy-paste that exists purely because two tools have no integration between them. The least interesting job in any business and one of the most expensive.

Every one of those has the same shape. The output is text or structured data, it is checkable before it leaves the building, and a mistake costs an hour rather than a reputation.

Why admin automation is the low risk bet

Three reasons, none of them complicated. It sits at the cheap end of the compute curve, where the same work costs a fraction of what a picture costs. A mistake gets caught by the person who reviews it, because the output is internal, reversible and checked before anyone outside sees it. A generated image is the opposite: public, unverifiable and permanent.

Nobody has ever felt deceived by an automated invoice reminder. That is the whole point, and it is why we build this half of the work and do not sell generated content. The return is real, and it does not come out of the trust a business spent years earning.

Where generated imagery still earns its place

This is not an argument that the tools are worthless. They are genuinely good at a few things, and pretending otherwise to make a cleaner argument would be its own kind of slop.

Internal mockups, where the point is to show a layout rather than a photograph. Storyboards. Concept exploration before a real shoot, when you are choosing a direction and nobody outside the room will ever see the frames. Throwaway visuals nobody could mistake for a photo.

The line is not hard to find.

It is whether a customer could reasonably believe a real person or a real place is in front of them. Inside that line, the tools save real time and cost almost nothing. Outside it, you are spending trust to save a few hundred dollars on a shoot.

How to decide what to automate first

A test you can run on Monday. Four questions, one point each.

Is the task repetitive, done the same way every time? Is the output checkable by a person before it matters? Does a mistake cost hours rather than reputation? Would a customer be unbothered to learn AI had done it? That last question is the one an AI use policy should already answer for your team.

Four yeses and it is worth automating. One yes and leave it alone.

Most businesses find three or four jobs clear the bar comfortably, and they are always the dull ones, which is the point. If you want a second opinion on which of yours clear it, an AI audit is exactly that conversation, and we are happy to give one. Plenty of the answers we hand back are that a process is too rare or too messy to pay for itself.

For the mechanics of doing it, how to automate a business process covers the mapping and the build. For the arithmetic, is AI automation worth it works through the numbers, and operations automation is what the finished thing looks like.