Andrew L. Urban.
The Labor government’s proposed Digital Duty of Care laws are in reality intended to override algorithms with Alborithms, algorithms operating under the authority, assumptions and ideological preferences of regulators and governments. Voters have a ‘duty to care’ about this potential threat to civil liberties (and the US has warned allies against imposing giant fines on tech giants).
The crucial transition is from “algorithmic recommendation” to “algorithmic permission”. This is where the issue becomes much more profound. Suppose Facebook’s algorithm decides: “Andrew is likely to engage with this story.”
Under the present system, the platform makes that prediction. Under the Digital Duty of Care, the government can require the platform to consider: “Is showing this category of material to Andrew reasonably foreseeable to cause harm?”
That sounds innocuous. But someone—or something—must make the classification. The system therefore acquires another layer:
Content → classification → risk assessment → regulatory rules → recommendation algorithm → user
Public discussion tends to focus on the last stage—the recommendation algorithm. But the politically consequential part may become the classification layer.
“Harm” is necessarily a normative concept. This is perhaps the strongest objection. Nobody seriously disputes that child sexual exploitation, grooming or material facilitating suicide can constitute serious harm. The difficulty begins further down the spectrum.
The government’s proposed framework encompasses harmful material and activity affecting young people and refers to categories including pornography, eating disorders, misogyny, glorification of crime and material associated with mental-health distress.
But consider:
Who determines whether something is misogynistic?
Or:
When does controversial discussion of sex differences become harmful misogyny?
Or:
When does discussion of suicide become encouragement of suicide?
Or:
When does political advocacy become extremist material?
Those are not purely technical questions. They involve judgment.
And once government establishes the regulatory categories, technology companies have a powerful incentive to err on the side of removal or suppression. That produces what might be called algorithmic precaution:
“We don’t know whether this is prohibited, so don’t show it.”
The danger is not necessarily censorship—it is invisible selection.
This is actually more subtle than conventional censorship.
Traditional censorship says:
You may not say X.
Algorithmic governance can say:
You may say X, but we won’t recommend it to anyone.
The distinction is enormous.
The speech technically remains available.
But if the platform’s recommendation system determines what millions of people encounter, demotion can be almost as consequential as prohibition.
That creates a potentially powerful form of soft censorship.
Nobody has banned the material.
Nobody has prosecuted its author.
Nobody has ordered the website closed.
It simply disappears from the information environment most people actually inhabit.
And that is where “Alborithms” becomes a useful political concept. You could define an Alborithm as:
An algorithm whose operation is shaped not merely by commercial optimisation but by politically determined concepts of acceptable behaviour, harm, safety and social responsibility.
And there is an important irony. The government’s justification is democratisation: Give people control over algorithms.
But simultaneously the state is acquiring considerably more influence over the conditions under which those algorithms operate.
The most important question is: who writes the rules?
This is where parliamentary accountability becomes crucial. The exposure draft gives the regulatory architecture considerable flexibility because technology and online harms evolve rapidly. The government does not want Parliament to have to amend legislation every time a new harmful digital phenomenon emerges. That is understandable.
But flexibility also means delegated power. And delegated power combined with automated enforcement creates a potentially formidable system:
Parliament → Minister → regulator → regulatory standard → platform compliance system → algorithmic enforcement → citizen
The citizen may never know precisely why particular material was suppressed, demoted or excluded. That is the real constitutional and democratic question.
The real test should therefore be neutrality.
A good Digital Duty of Care should regulate the mechanism of harm, not the political content of speech.
That distinction could be expressed as: Regulate the algorithm’s behaviour, not the algorithm’s conclusions.
For example, it is defensible to require:
- transparency about recommender systems;
- independent auditing;
- limits on targeting children;
- controls against addictive design;
- meaningful user choice;
- protection against grooming and exploitation;
- disclosure of recommendation criteria;
- independent research access;
- appeal mechanisms;
- explanations for significant moderation decisionsIt becomes considerably more dangerous when regulation starts determining:which political, cultural, religious or ideological viewpoints an algorithm should preferentially suppress or promote.
That is the point at which a Digital Duty of Care can become an ideological duty of conformity.
In short, the danger of Australia’s Digital Duty of Care is not that it will replace algorithms with artificial intelligence; it is that it may replace commercially programmed algorithms with regulatory “Alborithms”—automated systems whose definition of harm, safety and acceptable speech ultimately reflects judgments made by governments, regulators and politically appointed institutions.
That, in my view, is the genuinely serious civil-liberties issue raised by the proposal.