Two years ago, every keynote promised that artificial intelligence would write our emails, book our holidays and quietly reorganise our lives by Christmas. Sitting here in mid-2026, the reality is both more useful and more boring than the slide decks suggested. Our team has spent the last eighteen months using these tools for actual work — transcribing interviews, summarising court documents, editing product photography, debugging code at 11pm — and the gap between what the marketing claims and what we reach for daily has never been wider.
Priya, who heads up our AI desk, puts it this way: the genuinely useful features are the unglamorous ones. Nobody is filming a TikTok about meeting notes that finally make sense, but that is the bit that has changed how we work. The “agent that runs your life” demos still mostly fall over the moment you ask them to do anything load-bearing. So here is our honest stocktake of consumer AI in 2026 — what is worth paying for, what is worth ignoring, and what to watch for if you live in Australia.
What actually changed in 2025–26
The headline shift was that the frontier assistants — ChatGPT, Claude, Gemini and Copilot — stopped feeling like party tricks and started feeling like competent junior staff. Three concrete things moved the needle:
- Long context windows became normal. Pasting a 60-page PDF and asking sensible questions about it is now table stakes. A year ago it was a paid extra.
- On-device AI arrived properly. Apple Intelligence on the iPhone 16 and 17 lines, Galaxy AI on the S25 family, and Pixel AI on the Pixel 10 all do meaningful work locally — summarising notifications, cleaning up photos, transcribing calls — without round-tripping to a server.
- Voice modes finally stopped being awkward. Real-time, interruptible voice conversations with an assistant in your earbuds work well enough that we use them for hands-free dictation while driving or walking the dog.
What did not change: the underlying tendency of these models to invent things confidently. The hallucination problem is smaller than it was, but it has not disappeared, and any workflow that assumes the model is correct without a human glancing at the output is asking for trouble.
The features we actually use every day
If we audit our own usage logs — and Priya did, because she is that sort of person — a handful of features account for almost all of it. None of them are flashy.
- Summarisation. Long email threads, government consultation papers, ASX announcements, Slack backlogs. A two-paragraph summary at the top of a 40-message thread saves real minutes every day.
- Transcription. Whisper-class transcription is now accurate enough for Australian accents (finally) that we transcribe every interview and every internal meeting. Searchable transcripts have quietly replaced “I’ll remember that bit.”
- Image editing. Object removal, sky replacement, generative fill for product shots. What used to be a twenty-minute Photoshop job is now a thirty-second prompt. Magic Editor on Pixel and Clean Up on iPhone both work well enough that we no longer keep a designer on retainer for routine cleanup.
- Search and “explain this”. Perplexity, ChatGPT Search and Gemini’s grounded answers have replaced a real chunk of our Google usage for anything that needs synthesis across multiple sources. The citations are not perfect, but they are checkable.
- Code assistance. Cursor, Claude Code and Copilot have stopped being autocomplete and started being a competent pair programmer. Our developers reckon they ship roughly thirty per cent more in a day than they did pre-2024, with the caveat that the model still cheerfully invents APIs that do not exist.
- Translation and tone-shifting. Drafting a polite-but-firm email, translating a Mandarin product spec, rewriting a paragraph for a different audience. Small jobs, done in seconds, dozens of times a week.
Notice what is on that list and what is not. There is no “AI booked my flights” or “AI ran my business.” The useful stuff is assistive — it helps a human do a thing faster — not autonomous.
The features that are still mostly demo-ware
This is where we part company with the keynote energy. Several heavily marketed categories remain, to be polite, aspirational.
- Agents that “do things for you.” Booking restaurants, filling in forms, navigating websites on your behalf. We have tried every well-known one. They work for trivial cases and fall over on anything with a CAPTCHA, a login wall, a date picker or an unusual layout — which is to say, most of the real internet.
- “AI assistant runs your inbox.” Auto-drafted replies are fine for “thanks, see you Tuesday.” For anything that requires judgement, they still produce the corporate-bland sludge that everyone has learned to recognise and ignore.
- AI-generated video for serious use. Sora, Veo and Runway produce extraordinary short clips. Putting them into a real edit alongside real footage still looks uncanny in ways viewers clock immediately.
- “AI search” on smart glasses. The hardware has improved dramatically — if you are curious about where the form factor is heading, we covered the current state of play in our piece on augmented reality glasses — but the assistant layer on top is still mostly novelty.
- On-device “personal context” features. The pitch that your phone’s AI will quietly read all your messages, calendar and photos to give you eerily relevant suggestions is technically working. In practice, the suggestions are still about as useful as the ones Google Now offered in 2015.
None of this means the categories are dead. It means that paying a subscription today for the promise of tomorrow’s agent is paying for a beta test.
The Australian privacy picture
For Australian users, the privacy questions around all this are not academic. A few things to keep in your head:
- Data residency varies wildly. Apple Intelligence’s Private Cloud Compute keeps requests inside Apple’s encrypted infrastructure. ChatGPT and Gemini, by default, may process queries in US data centres. Microsoft Copilot for business customers can be pinned to Australian regions; the consumer version cannot.
- Training opt-outs are inconsistent. OpenAI, Anthropic and Google all offer ways to stop your conversations being used to train future models, but the toggles are buried in different places and reset surprisingly often after updates.
- The Privacy Act reforms are landing. The first tranche of changes from the 2024 review is now in force, with the second tranche progressing through Parliament. The Office of the Australian Information Commissioner has published useful plain-English guidance on how generative AI fits the new rules — worth a read if you handle customer data.
- Online safety is a separate question. The eSafety Commissioner has been increasingly active on AI-generated material, particularly around deepfakes, with industry codes that affect what consumer apps are allowed to ship in Australia.
The practical takeaway: assume anything you type into a free consumer AI is being seen by the provider, even if it is not being used to train. Treat it like an email you would not mind a stranger reading.
The new attack surface
AI has also given scammers a serious upgrade, and that is the bit we cover most often on this site. Voice-cloning scams targeting elderly Australians, AI-generated invoice fraud at small businesses, and convincing phishing emails free of the broken English that used to be a tell — all of it is now cheap and scalable.
A few defensive habits worth adopting:
- Agree a family or business safe word for any unexpected “I need money urgently” call, regardless of whose voice it sounds like.
- Be sceptical of polished, perfectly written emails from suppliers asking you to update bank details. Polish is no longer a trust signal.
- Hover over every link. Our guide to spotting a scam website still holds up — AI has not changed the underlying tells, it has just made the surrounding pitch more convincing.
- If your phone is behaving oddly after clicking something suspect, our walk-through on whether your phone has been hacked covers the signs to look for.
What we would subscribe to in 2026
Priya’s rule of thumb is simple: pay for the assistant you actually open more than once a day, and ignore the rest. Our team’s honest spending list, after culling subscriptions over the past year:
- Worth it: one frontier chat assistant (ChatGPT Plus, Claude Pro or Gemini Advanced — pick one, they are largely interchangeable for consumer use); a transcription tool with good Australian-accent support; a coding assistant if you write code for a living.
- Probably free is enough: on-device features that ship with your phone, basic image cleanup, voice mode for casual queries.
- Skip for now: standalone “AI agent” subscriptions, AI-only hardware (the AI pin and rabbit-style devices are still searching for a job), AI-generated video services unless you are a professional creator.
The other thing worth doing: cancel ruthlessly. Most of these services launched with annual plans and aggressive discounts. The market is moving so fast that a tool that was best-in-class last September may be middle-of-the-pack now. We re-audit our subscriptions every quarter and have not regretted dropping any of them.
Final thoughts
The honest summary of consumer AI in 2026 is that it has become genuinely, quietly useful for a narrow set of tasks — summarisation, transcription, image cleanup, search, code — and remains overpromised for almost everything else. The keynotes will keep showing demos of assistants running your life. The real productivity gain is sitting in the corner of your inbox, turning a 40-message thread into two paragraphs you can actually read.
Our advice for the next twelve months: pick one assistant and learn it properly rather than dabbling in five, treat anything an AI tells you as a draft rather than an answer, mind what you type into free services, and keep a healthy scepticism for any product whose pitch begins with the word “autonomous.” The useful stuff is here. The magic is not — and that is fine.
