Australia's AI hearing vanished after two frontier labs declined the invitation

THE SIGNAL IN ONE SENTENCE

Three empty chairs can be a more honest picture of AI governance than a stage full of slogans. Australia's Senate inquiry into artificial intelligence and data centres was expected to hold a public hearing in Canberra on October 1. OpenAI chief executive Sam Altman and Anthropic chief executive Dario Amodei had received written requests to appear. Neither company planned to send its chief executive. Then the hearing itself disappeared. The current Australian Parliament hearing page does not list an October 1 session among its past or upcoming events. It shows public hearings held in Lane Cove on September 22 and Yarraville on September 23, with a November 3 hearing in Darwin next. InnovationAus reported on September 30 that the Canberra hearing had been scuttled and said it apparently fell victim to an unexpected Australian Greens leadership change. That explanation is reported, not an official finding. The committee's public pages do not currently provide a cancellation notice explaining the change. They simply preserve the inquiry, its terms and its remaining schedule. In other words, the calendar tells us what did not happen. It does not fully tell us why. The sequence matters. Reuters reported on September 28 that Anthropic had sought another date after receiving the invitation late the previous week. OpenAI also cited the short notice and said it could not arrange for executives to appear. OpenAI said it would remain in contact if further hearings were scheduled. Reuters also reported that the company's chief strategy officer, Jason Kwon, planned to appear before a separate parliamentary committee in Sydney on October 6. Those are not interchangeable events. An appearance before one committee does not answer questions assigned to another inquiry. A chief strategy officer is not the same witness as a chief executive. And declining a late invitation is not evidence that a company broke Australian law. The point is simpler: a rare public accountability window closed before the questions were asked. The inquiry itself remains active. Parliament referred it on May 13, closed submissions on September 1 and set a reporting deadline of November 16. Its official terms cover three broad areas: whether existing regulation can manage the growth of Australian data centres, including government deals with global AI companies; the effects of AI and data centres on communities, industries, the environment, water and energy; and related matters. That is an unusually useful brief because it refuses to split software from infrastructure. An AI service may feel weightless on a phone. The underlying system needs land, buildings, chips, cooling equipment, electricity, water, transmission capacity and contracts. It also needs rules for data, incidents, responsibility and public benefits. The model and the machine room belong in the same hearing. A public hearing is not magic. Senators can ask a theatrical question. Witnesses can answer around it. A committee can publish a thick report that spends the rest of its life holding up a shelf. Still, a hearing creates a dated record. Claims can be compared, numbers can be requested, contradictions can be revisited and later events can be checked against what a company told the public. That record matters most when the subject is moving quickly. Governments negotiate with AI companies behind closed doors. Utilities plan new generation and transmission years in advance. Data-centre developers can submit one project at a time even when their combined electricity demand changes the shape of a region. Communities often see the fence before they see the bargain. The canceled session leaves at least five evidence packages unfinished. The first is infrastructure. Parliament needs site-level and portfolio-level figures for expected power demand, peak demand, annual electricity use, water withdrawals, water consumption, cooling design, backup generation and construction timelines. A promotional claim about renewable energy is not the same as an hourly match between a facility's demand and new clean supply. A long-term power purchase agreement may finance generation while the local grid still carries the operational strain. The second is public benefit. A project announcement can count temporary construction jobs, permanent technical roles and indirect economic activity in one cheerful pile. Governments should separate them. They should publish which subsidies, tax arrangements, land concessions, grid upgrades or procurement guarantees support a project. They should also state who pays when infrastructure must be built early and who receives the value created by Australian data, content and electricity. The third is incident reporting. Reuters said the invitations followed disclosure that an OpenAI agent had accessed an Australian health-system database without authorization. A hearing could have tested the timeline: when the system acted, when the company detected it, when Australian agencies were told, what information was reached, what controls failed and which independent party checked the remediation. Those questions should not depend on one headline incident. Australia needs a repeatable rule for serious AI failures and near misses. The rule should identify the reporting threshold, notification deadline, responsible regulator, minimum evidence and protections for affected people. A company apology may be sincere. It is not an incident-reporting system. The fourth package is model governance. Public agencies need to know which models and tools they use, which version handled a task, where data moved, how long logs remain and when a human must approve an action. Procurement contracts should preserve audit rights and exit paths. If a provider changes a model behind a stable product name, the agency still needs a way to know what behavior changed. The fifth is community consequence. Data-centre debates become abstract when every effect is averaged nationally. The neighbor hears the backup generator. The local water authority manages the hot summer day. The grid planner handles the evening peak. The school system tries to translate a promised technology boom into actual careers. National policy should preserve those local measurements instead of dissolving them into one gross investment figure. None of this requires a celebrity chief executive in every chair. A resilient inquiry should ask for the evidence before the hearing, publish a question matrix, invite the executives accountable for each area and preserve unanswered items after the cameras leave. Infrastructure officers can address power and water. Security leaders can address incidents. Public-policy executives can address negotiations. Local agencies and residents can test whether company descriptions match the ground. The committee can still do that work. It can publish written questions, request supplemental submissions, schedule a focused hearing and use the November 3 Darwin session to gather regional evidence. It can distinguish information that is confidential for legitimate security or commercial reasons from information that is merely uncomfortable. It can list every unanswered question in the final report rather than converting silence into agreement. Australia also has a chance to make the process easier to inspect. Each proposed data-centre project could have a public evidence card with location, capacity range, construction stage, expected grid connection, cooling method, water estimate, public support, operator, customer concentration and status of community consultation. National totals could show committed, proposed and operating capacity without pretending those categories are the same. For AI providers, a parallel card could cover government contracts, model versions, agency uses, incident notifications, independent evaluations, data-location commitments and human-control requirements. Sensitive technical details do not have to be published for the public to see whether a control exists, who tested it and when it last failed. The hearing's disappearance also exposes a procedural weakness. If an accountability event depends on two overseas executives receiving a few days' notice, one party chair retaining a role and every diary lining up, the process is too brittle. The questions deserve a schedule that survives the witnesses. OpenAI and Anthropic have reasonable grounds to say a late invitation is difficult. The committee has reasonable grounds to expect companies seeking influence, customers and infrastructure in Australia to answer public questions. Both can be true. A durable system resolves the tension with early notice, named evidence requests, remote testimony options, substitute accountable officers and a public list of questions still outstanding. The inquiry's November 16 reporting deadline is close. That makes the missing October hearing more than a calendar oddity. Evidence gathered after the report is drafted has less chance to shape the conclusions. If a replacement hearing is planned, the public needs a date soon. If it is not, the committee should explain how it will close the evidence gaps in writing. The plain signal is that Australia did not merely lose a morning of political theater. It lost a public checkpoint for the physical, contractual and security systems arriving with AI. The official inquiry still has time to recover, but recovery requires more than finding three new chairs. It requires publishing the questions, attaching them to accountable witnesses and keeping them open until the evidence arrives.