top of page
Search

AI: The Drug We Didn't Know We Were Taking

AI: the drug we didn't know we were taking

The dose matters. The side-effects matter. But what happens when something goes wrong?

AI may be a remarkable medicine. The problem is that we are prescribing it to everyone, for everything, in ever-increasing doses, before we fully understand the side-effects.

Imagine a pharmaceutical company announcing a revolutionary new drug. It promises to make us more productive, improve medical research, reduce costs, solve complex problems, help manage climate change and perhaps transform society.

There is just one complication. The company cannot tell us precisely what the long-term side-effects will be. Its developers and users cannot always fully explain how a particular AI system arrived at an output. The environmental cost of the infrastructure required to provide it is increasing rapidly. The dosage being consumed is growing at extraordinary speed. Children are beginning to use it. Businesses are becoming dependent on it. Governments are building infrastructure around it.

And rather than conducting controlled trials before widespread adoption, the world is effectively conducting the trial on itself.

Would we approve such a pharmaceutical product? Probably not.

Artificial intelligence is not literally a drug, and using an AI assistant is not the same as developing a physiological addiction to an opioid. But as a way of thinking about benefit, dosage, dependence, side-effects, externalities and regulation, the pharmaceutical comparison is surprisingly useful.

Perhaps we have been asking the wrong question. Instead of continually asking whether AI is good or bad, we should be asking: What is the right dose? Who should be taking it? For what purpose? What are the side-effects? Who pays for them? And what happens when something goes wrong?

Powerful drugs can do wonderful things

Powerful drugs can do wonderful things

Modern medicine would be almost unimaginable without powerful pharmaceuticals. Morphine can relieve extraordinary pain. Antibiotics have saved countless lives. Benzodiazepines have legitimate clinical uses. Chemotherapy deliberately exposes the body to toxic substances because, in the right circumstances, the potential benefit outweighs the harm.

We do not describe these medicines simply as good or bad. We ask whether their use is justified. We consider the patient, the condition, the dose, the duration, interactions, adverse effects and alternatives.

Artificial intelligence deserves the same mature discussion. AI can help researchers analyse huge datasets, improve accessibility, accelerate scientific discovery, optimise complex systems and assist people in work that would otherwise be difficult or time-consuming. There are applications where substantial computing resources may produce enormous social value.

But the existence of those benefits does not mean every use of AI is automatically beneficial. Medicine does not work that way. Neither should technology.

When something useful becomes something we depend upon

For example: let's look at opioids which provide one of the clearest parallels. Medicines such as morphine can be extremely effective for pain. But prolonged use, misuse and unsupervised use can lead to dependence and other harms. The World Health Organization notes that opioid dependence can include increased tolerance, withdrawal symptoms and an increasing priority given to opioid use despite negative consequences.

There is an interesting parallel with AI. Only a few years ago, asking a computer to draft a coherent report or create a credible image seemed extraordinary. Today AI is routinely used to write emails, summarise meetings, create advertisements, answer customers, analyse documents, write software, generate images, prepare presentations and make recommendations.

Each time AI makes something easier, there is an incentive to use more of it. The extraordinary becomes ordinary. Once it becomes ordinary, removing it becomes difficult.

That is not necessarily clinical addiction, but it can become technological dependence. The individual becomes dependent. Then the business becomes dependent. Then an industry becomes dependent. Eventually society builds infrastructure around that dependence.

Much like tolerance to a drug, yesterday's extraordinary dose becomes today's normal dose and tomorrow we want more.

Of course, there are many other examples of pharmaceuticals that have caused major problems when misused or overused.

What happens when AI gets it wrong?

Humans make mistakes. AI systems also make mistakes. The important difference is not that machines are uniquely fallible; it is that automation can multiply a mistake at extraordinary speed.

A person can make one incorrect decision. An automated system can potentially apply the same flawed assumption to thousands or millions of people before anyone notices.

Imagine an AI system incorrectly screening job applicants, prioritising medical cases, assessing insurance claims, flagging fraud, generating software code, interpreting intelligence, allocating public services or controlling infrastructure. The failure may originate in bad data, model error, a poorly written instruction, an incorrect assumption, cyberattack or simply a problem the designers never anticipated.

The danger grows with dependence. If an organisation has removed the people, processes and skills that once performed the task, what is its fallback when the AI system is wrong or unavailable?

This suggests that every serious AI deployment should have the technological equivalent of pharmacovigilance: monitoring after deployment, adverse-event reporting, human escalation, rollback procedures and a genuine ability to withdraw the system when harm emerges.


Our dominant AI philosophy often appears to be: deploy it, scale it, discover the consequences, then regulate the problems later.

That approach deserves scrutiny when AI is being embedded in education, medicine, finance, employment, media, government, defence and critical infrastructure. Some errors can be corrected. Others can spread too quickly, affect too many people or become too deeply embedded to reverse easily.


Australia's AI dose

The environmental cost is physical, not virtual


Australias AI dose The environmental cost is physical, not virtual

AI can appear almost weightless. Ask a question on a phone and words appear seconds later. There is no exhaust pipe, fuel tank or cooling tower beside the device, so it is easy to imagine that AI exists in an environmentally neutral thing called "The cloud".

It does not. The cloud is industrial infrastructure.

Behind AI services are data centres, servers, high-performance processors, substations, transmission networks, cooling equipment, backup systems, buildings, water systems, mined materials and enormous amounts of capital.

AI is not the only reason data centres are growing. Cloud computing, streaming, telecommunications, finance, government services and ordinary internet activity all require them. But AI is materially changing the scale and power density of the infrastructure being proposed.


Electricity: the dose is becoming significant


AEMO's 2026 Electricity Statement of Opportunities provides a useful measure of the scale. It forecasts data-centre electricity consumption in the National Electricity Market increasing from approximately 5 terawatt-hours in 2025-26 to 34 TWh by 2035-36. That would increase the sector's share of electricity supplied through the grid from around 3% to approximately 13%.

That is not a marginal change.

Australia is simultaneously electrifying transport and industry, replacing ageing coal generation, building renewable-energy zones, adding storage and constructing new transmission while maintaining reliability through extreme weather. Into that transition we are adding one of the fastest-growing new sources of electricity demand.

Saying that data centres can use renewable energy is not the end of the discussion. New renewable generation still requires land, transmission, storage, connection infrastructure and capital. If data-centre demand requires additional generation or network capacity, that infrastructure must actually be built and someone must pay for it.


The correct question is not simply 'Does the data centre buy renewable electricity?' It is 'What additional demand has been created, and what additional generation, storage and network capacity has been created to supply it?'

Water: another ingredient hidden from the user


Water is another part of the prescription. Depending on their design and cooling technology, data centres can require substantial water resources.

In its September 2025 determination for Sydney Water, IPART reported that Sydney Water estimated the water needs of data centres may be up to 250 megalitres a day by 2035. IPART also stressed the considerable uncertainty around the scale, number and timing of future facilities. The figure should therefore be treated as a scenario, not as a forecast of inevitable consumption.

But that uncertainty is a reason to plan early, not a reason to ignore the issue. Australia is not a country where water can sensibly be treated as an unlimited industrial input.

We already expect mining, agriculture, manufacturing and urban development to account for their water needs. A digital product should not receive an exemption simply because the resource consumption occurs out of sight.


NSW has started asking who pays


In August 2026 the NSW Government released a Data Centre Policy Framework and Guidelines aimed at managing energy, water and environmental impacts while facilitating investment. The principles include world-class environmental performance, no net cost to consumers and communities, and funding additional supplies of water and energy.

The scale of the pipeline helps explain why this matters. Infrastructure NSW says NSW has more than 60 data centres operating or under construction, with a further 19 State Significant Development projects in the pipeline valued at $50.3 billion.

The framework is accompanied by proposed reforms to electricity network connection and cost recovery, and an IPART review of the water pricing framework for data centres. As of September 2026, those processes are still under consultation and development.

This is significant. When governments need new rules to determine who pays for the electricity and water infrastructure supporting data centres, AI and cloud infrastructure are no longer merely technology stories. They are energy, water, planning, environmental and economic policy.

AI can help the environment - but that doesn't cancel its footprint

AI may help improve weather and climate modelling, optimise energy systems, identify inefficiencies and support scientific discovery. Those benefits should be recognised.

But they do not logically cancel the environmental costs of AI itself.

The two questions must be separated. First: can a particular AI application help reduce emissions or environmental harm? Second: what electricity, water, land, infrastructure and materials are required to provide that AI? A positive answer to the first question does not make the second question disappear.

Imagine a pharmaceutical manufacturer saying that the environmental impacts of producing one medicine should be ignored because the company also makes medicines that improve health. We would reject the logic immediately.

AI deserves the same accounting discipline.

Where is the AI environmental prescription label?

One of the first things we do when collecting a prescription is read the label. It tells us what we are taking, how much to take, and the warnings and precautions we need to know.

There is a strange asymmetry in the way consumers are informed. When Australians buy many appliances they can see energy-efficiency information. Packaged food carries nutritional information. Prescription medicines arrive with dosage instructions, contraindications and warnings.

But when we use an AI service, we generally know almost nothing about the environmental cost of the service. We do not know where the request was processed, the electricity or water attributable to the service, the emissions associated with the location, or whether a much smaller model could have achieved the same result.



The dose makes the difference

There is an old toxicological principle often simplified to: "the dose makes the poison".

It may be equally useful when thinking about artificial intelligence.

Using substantial computing resources to help discover a cancer treatment or predict catastrophic bushfire behaviour may produce enormous social benefit. Using similar technology to generate millions of disposable advertisements, rewrite trivial emails or create synthetic content designed only to fill social-media feeds is a very different proposition.

Both are called AI. Their social value is not equivalent. Their environmental justification should not automatically be treated as equivalent either.

The commercial market, however, generally rewards increased consumption. More AI usage produces more subscriptions, more cloud consumption, more processors and more infrastructure. The economic incentive is therefore not necessarily to determine the appropriate dose. It is often to increase it.

AI Stewardship

AI stewardship: prescribe it rather than simply consume it
AI stewardship: prescribe it rather than simply consume it

Antibiotic stewardship does not mean banning antibiotics. It means using them appropriately, choosing the right treatment and avoiding unnecessary use.

We could adopt a similar idea for artificial intelligence:

AI Stewardship.

  • Use AI where its measurable benefit genuinely outweighs its financial, environmental and social costs.

  • Use smaller models, conventional software or human judgement where they can perform the task adequately.

  • Require stronger scrutiny, testing and human oversight for high-risk applications.

  • Measure electricity, water and infrastructure impacts instead of hiding them inside the cloud.

  • Maintain human capability and fallback processes so organisations can function when AI fails or becomes unavailable.

  • Monitor deployed systems and report serious adverse events, just as medicine monitors unexpected harms after approval.

  • Require a practical recall mechanism: the ability to stop, roll back or replace an AI system when something goes wrong.

Most importantly, do not use AI simply because AI is available. Doctors do not prescribe morphine because there happens to be some in the cupboard.

What should success look like?

We do not measure the success of medicine by counting how many tablets society consumes. We measure outcomes: lives improved, pain reduced, disease prevented and harm avoided.

Perhaps we should stop measuring AI progress by the number of tokens processed, GPUs installed, data centres built or dollars invested.

We should ask what the AI actually improved. What resources did it consume? What risks did it create? Who received the benefit? Who paid the cost? What happens if it fails? And could we have achieved the same result with a smaller dose?

The warning signs may be ordinary, not dramatic

The greatest risk from artificial intelligence may not be a science-fiction machine suddenly deciding to turn against humanity.

It may be something much more ordinary.

We keep increasing the dose because every individual use feels convenient and relatively harmless.

Useful becomes normal. Normal becomes expected. Expected becomes necessary. Necessary becomes dependence.

Meanwhile the factories producing the dose become larger. Electricity demand rises. Water demand rises. Infrastructure expands. Human skills quietly atrophy. Automated decisions become harder to challenge. And the ability to operate without the technology slowly disappears.

Only then might we discover what the cumulative side-effects really were.

The question is no longer whether we should use AI. We already do. The question is how much AI is enough, and whether we will recognise the warning signs before we overdose.

Maybe we already have.


References:


Author: John Debrincat FACS MAICD

ShapedLogic Blog 2026


Authors Note: ChatGPT was used to create the images used in this article. Gemini and ChatGPT were used to assist in fact checking and material searches. The core article and any poor grammar was entirely my doing.

Comments

Join ShapedLogic News

Thanks for joining!

PO Box 635, Gladesville, NSW 1675

  • Facebook
  • Twitter
  • LinkedIn

©2024 by ShapedLogic ABN: 37 815 761 246 - Proudly Australian - Privacy Policy

bottom of page