I held the button on my phone and said a few words, they were not a question for an LLM, they were search terms. I wanted the results page, my words, passed to an index, ten links back, and I would pick the one I wanted. What I got instead was Gemini, thinking, then a paragraph of synthesised text answering a question I had not asked, with the thing I actually wanted somewhere underneath it.
That interaction is no longer the exception. It is the default, and the default is spreading. Google announced in March 2025 that Gemini would replace Google Assistant on Android, and when the transition completes in 2026, Assistant will no longer be available on Android phones and tablets at all. Google's own support pages state that on the Pixel 9 and later, Gemini comes as the default assistant. Microsoft has done the same thing to search itself: Bing once had a setting to disable Copilot answers, and that setting has been removed. Google's AI Overviews became default-on for most accounts by early 2026, and the opt-out that existed in Search Labs during the rollout is gone.
The pattern is familiar to anyone who has watched a vendor chase a metric. A capable, expensive tool is inserted into a path where a cheap, boring tool was doing the job, because usage of the expensive tool is the number being reported upstairs. What follows is what always follows.
The energy argument is weaker than you would expect
I will concede the popular version of the environmental complaint up front, because it does not survive contact with the numbers.
The line you have heard is that an AI query burns ten times the energy of a Google search. That figure comes from a 2023 estimate of roughly 3 watt-hours per ChatGPT query, set against a 0.3 watt-hour per-search figure Google published in 2009. Both halves have aged. Google's own technical paper, published in August 2025, puts the median Gemini Apps text prompt at 0.24 watt-hours, and claims the median prompt's energy footprint fell 33 times in a single year. Epoch AI independently estimated a typical GPT-4o query at around 0.3 watt-hours. The International Energy Agency went further: if every conventional internet search on Earth were performed as a simple AI text query, the total would come to less than 4 terawatt-hours a year, under one percent of what data centres already consume.
So if the argument were only about the marginal watt-hours of my one voice query, it would be a weak argument. It is not only about that, and the honest numbers make the real problem clearer, not smaller.
The result is worse
Start with the part that needs no measurement equipment: the tool fails at the job it displaced.
Google Assistant answered fast because it was simple. Speech recognition, a lookup or a scripted action, done. Gemini routes everything through a large language model first. One writer who switched back described requests taking four to five seconds longer than Assistant, with the answer arriving wrapped in paragraphs of preamble, and community forums carry many reports in the same shape. On Google Home the regression was severe enough that threads filled with reports of lights taking twenty to thirty seconds to turn on, and 'it's almost worthless' became the representative quote. Google's response was to announce it had reduced smart home latency by 40 percent, which is the polite way of confirming the regression was real.
This is the predictable failure. A voice query on a phone is overwhelmingly a retrieval task. The person already knows what they want. They need information, not interpretation. Putting a reasoning engine in front of a retrieval task adds latency, adds verbosity, and adds a layer that can misread intent, in exchange for capabilities the task never called for. Assistant was a doorway. Gemini is a concierge who insists on describing the building.
One query is no longer one query
Here is the part the per-query energy figures quietly leave out.
When Gemini or AI Mode handles a query, it does not run one search in place of my one search. Google's own description of the technique, which it calls query fan-out, is that the system breaks the question into subtopics and issues a multitude of queries simultaneously on your behalf. Google's own blog puts it plainly: AI Mode is basically doing a dozen searches for you in the time it takes to do one. The Deep Search variant, per Google's product lead for Search, can issue hundreds. So the honest accounting for my simple voice query is not one model inference versus one search. It is one model inference, plus a dozen or more hidden searches, plus a synthesis pass, to produce a paragraph answering a question I could have answered myself from the first result.
Multiply that by the scale of the substitution and the aggregate picture stops being abstract. Data centres consumed around 415 terawatt-hours in 2024, about 1.5 percent of global electricity. The IEA projects that to more than double to around 945 terawatt-hours by 2030, with AI as the most important driver. Google's reassuring 0.24 watt-hour figure covers the median text prompt only. It excludes voice, excludes images, and excludes the reasoning-heavy models. One independent estimate, from the University of Rhode Island's AI lab and first reported by The Guardian, put a medium-length GPT-5 response at an average of just over 18 watt-hours. That is an estimate built from response times and hardware assumptions, not a measurement, and it has been challenged as high. The direction of it has not been. The per-query number got smaller. The number of queries, and the machinery behind each one, got much larger.
The off switch is gone
None of this would matter much if it were a choice. It is not.
There is no setting to turn off AI Overviews. There is no setting to restore Copilot-free Bing results. When the Assistant transition completes there will be no Assistant to fall back to. What exists instead is a folklore of workarounds: a Web filter buried in the results tabs, a udm=14 URL parameter you can wire into a custom search engine, the trick of appending '-ai' to a query. Ordinary users will never find these, which is the point. When a feature is good, you market it. When a feature is a metric, you remove the exit.
Good engineering routes a task to the cheapest tool that solves it. That principle holds at every scale, from a shell script to a national grid, and the industry currently making the largest infrastructure investments in history is applying it in reverse: routing the world's simplest, highest-volume task through its most expensive machinery, because inference volume is the number in the board pack. The models are genuinely capable. The datacentre buildout is genuinely enormous. And the task, most of the time, was a search.
I fix DNS records for a living. I know what it looks like when nobody asks whether the task needed the tool.