Executive Summary
Striking Blind uses fiction to consider the opportunities and challenges of military Artificial Intelligence (AI) and autonomous systems. Set in 2040, it describes the ADF’s use of MANDELA, the Machine Analysis Decision Learning Algorithm, a US-manufactured intelligence fusion, analysis and decision support tool inspired by AI systems currently in development. MANDELA provided the ADF with decision superiority during a stability operation in the Philippines, yielding invaluable improvements to decision speed, accuracy, and creativity. It reduced the risk to coalition ground forces through timely threat engagement and minimised the chance of civilian casualties through accurate collateral damage analysis.
However, the ADF’s employment of MANDELA also raised challenges. The ADF was pressured to field the system in response to Indo-Pacific competition, US interoperability requirements and a responsibility to minimise risk to combatants and civilians. Due to intellectual property rights and US arms export restrictions, MANDELA’s decision-making algorithms remained opaque to its human operators. Comprehensive test and validation of MANDELA proved difficult due to the vast number of rapidly changing intelligence inputs and the variability of its operating environment. Finally, its human operators became conditioned to trust it; after receiving hundreds of successful strike recommendations, they lost the willingness to challenge it.
With foresight, the ADF can address these challenges now. First, the ADF should closely collaborate with partner governments and industries to oversee the design of ADF AI systems and ensure their operation is continuously supported throughout their service. Second, the ADF must take a whole-of-system approach to certification and validation, as an AI’s reliability will be impacted by each component that interfaces with it: sensors, weapons, databases and networks. Finally, the ADF should consider a policy framework that recognises the influence AI systems may gain over human decision-makers. For example, when should these decision-makers be held accountable for accepting a poor AI recommendation, and under what circumstances should that accountability be held elsewhere?
By maintaining robust oversight of the ADF’s integrated weapon system, the ADF can benefit from AI and autonomous systems while addressing their complex challenges. This oversight will ensure the ADF avoids the human costs illustrated in Striking Blind.
STRIKING BLIND
15:23 04 Feb 2040, 300kts and 27000ft over the Philippines
From the rear cockpit of her F/A-18F Block IV Super Hornet, Jazz peered down the right side of the canopy over Vector’s shoulder; they were approaching the air-air refuelling drogue trailing behind the huge RAAF KC-30. She clenched the canopy grab rails tightly – knuckles white – subconsciously holding her breath and straining up against her lap straps to get a clear view. Vector was a good pilot, but air-air refuelling always unnerved her a little. She felt Vector’s control inputs twitch the aircraft as he fought to keep their probe aligned with the moving basket.
She stared at it intently, three feet to go… two feet to go…. The basket leapt out of the way! She shut her eyes tightly as the aircraft yawed and jerked after it. The probe hit the basket with a clunk and ground into the centre. Jazz exhaled in relief; she relaxed her grip on the grab rails and settled back down into her seat. It wasn’t Vector’s best plug, but she knew not to comment on it until they were safely on the ground – it was one of their unwritten rules.
Tension leaving her, she adjusted the long blonde plait tucked into the back of her flight suit. ‘Did I tell you Lucy started school yesterday?’ she asked Vector through the intercom. ‘Mike sent me some videos.’
‘Wow, that’s amazing! Bet you’re looking forward to getting home,’ he replied, picturing her daughter and making a slight correction to hold them in position. ‘How old is she now? Six?’
‘Five,’ Jazz said with a smile. ‘She’s growing up so quickly.’ She looked down at the large display in front of her. Their aircraft’s small blue position icon was 80 nautical miles south of Palawan Island in the Philippines. She scrolled the multi-coloured map towards the large island, looking for their wingman, callsign Cannon 2.
Vector saw her cursors move on his front cockpit display. ‘What are they up to?’ he asked.
‘Looks like they’re still escorting the patrol; it’s just south of Puerto Princesa now. Some neutral tracks around, but no enemy contacts.’ A few days ago, the Armed Forces of the Philippines (AFP) had recaptured Puerto Princesa from the separatists. This special operations patrol was part of a larger effort to consolidate their position. Their RAAF formation, callsign Cannon, was assigned to protect it.
Jazz watched Cannon 2’s green icon orbit overhead the patrol, arcing a large 10-nautical mile circle around it. Four other green circles represented the autonomous drones that accompanied them. These were small, uncrewed fighters that could be controlled from a headquarters or handed to an airborne platform. Today they were assigned to their formation to assist with Close Air Support. The engineers called them Loyal Wingmen; the flyers called them Bots.[i]
Jazz clicked opened a menu that displayed the tasks Cannon 2 had assigned to the drones. She noted two were tasked with Electronic Warfare, flying large L-shaped patterns, collecting, locating and selectively jamming emissions. The other two were set to Forward-Looking Infrared (FLIR) overwatch, scanning the patrol’s immediate vicinity using their optical sensors.[ii] Her display told her that each Bot still had most of its bombs remaining.
She looked out the left side of her canopy in their direction. Green boxes projected on her helmet’s visor overlaid their location. They were too far away to be seen visually; all she could see was an empty blue sky. Below, a layer of scattered cloud stretched across Palawan Island.
‘That’s top-off, Jazz,’ Vector announced. Jazz peered over his shoulder. Their aircraft was backing away from the basket, the probe disconnecting with a pop. ‘Are you ready to switch out with Cannon 2?’ he asked.
‘Got it,’ Jazz replied. Vector slid their jet out from behind the tanker and accelerated ahead, pressing Jazz back against her seat. She transmitted Cannon 2 a pre-programmed digital message: Cannon 1’s departing the tanker, Cannon 2 request hand-off.
Cannon 2’s acknowledgement message appeared on Jazz’s display. The four Bots reassigned themselves to her, their icons changing from green to blue. She watched Cannon 2’s icon turn towards the tanker; they’d be getting their last gas and heading home. She stretched the tight muscles in her neck—lucky for them, she thought, they’ll make it back before sunset.
She zoomed her map in, turning her attention to the patrol. Yellow tracks appeared 800 metres south of their position. Two at first, then several more. ‘Vector, we’re getting more tracks to the south,’ she said. Jazz felt the thrust increase as Vector accelerated them towards the patrol still 50 nautical miles away.
Jazz toggled her cursors between the tracks, bringing up their ID information; they were personnel. The contributor listed against them was LW-FLIR, meaning that at least one of the Bots had identified and classified them with its optics. She opened one of their sensor feeds. It showed dense trees with red boxes marking the location of the people it had detected. Occasionally there was a flick of movement through gaps in the foliage; you would overlook it unless you knew it was there. It was these changing pixels the bots were using for their detections.
The radio crackled, and the Australian Joint Terminal Attack Controller’s (JTAC) gruff voice came over the air: ‘Cannon 1, this is Havoc. We have reports of possible enemy activity to their south.’
Jazz confirmed the JTAC was aware of the Bots’ tracks—he was. The JTACs usually had better information than her. They sat in the headquarters north of Puerto Princesa with the commanders and their decision assistant, MANDELA. MANDELA was a room-sized artificial intelligence system. It analysed military and open-source information to identify targets and recommend their engagement, considering proportionality, distinction, necessity, and national targeting restrictions.[iii] It could direct engagements automatically, which was essential in high-intensity warfare; however, for this operation, it was limited to just providing recommendations—a commander would need to make the final decision.[iv] Of course, its recommendations were subjective with a degree of uncertainty, but humans were subjective too —in any case, MANDELA had proved itself more reliable; Jazz had never seen it make a mistake.
Cannon 1 arrived overhead of the patrol, and Vector wheeled the jet around to the left. Jazz looked down between gaps in the scattered cloud. The terrain was covered in dense green trees casting long shadows over the occasional patch of grassland. She couldn’t see any people. She returned to her map display, hovering her cursors over each of the tracks. A few of them had started to read LW-FLIR, GFR, indicating a friendly Ground Force Report was now correlated with the Bots’ tracks. They were closing rapidly with the patrol, now 500 metres away. One of the tracks turned red.
‘Woah, we got a hostile!’ Jazz said, snapping her cursors to the red track. The previous contributors were replaced with a single word—MANDELA, indicating the AI had recommended engaging. Jazz shifted in her seat expectantly, leaning against her harness towards the display.
‘Cannon 1, this is Havoc,’ the JTAC said. “Enemy forces are rapidly approaching the patrol. You are cleared to engage MANDELA tracks, ROE 556, Ground Force Commander Initials Mike Papa.’
‘Cannon 1, copied,’ Jazz replied. ‘Cleared to engage.’ More tracks appeared, each representing an enemy combatant, and in an outwards wave, they turned red.
‘How do you want to do this?’ Vector asked.
‘Bot 3 and 4, airburst, IR guidance, auto,’ Jazz replied while quickly selecting the menu options. ‘Hold us out to the west and target spitters.’
‘Got it! Hurry—they’re 400 metres out and closing!’ He banked hard right, throwing Jazz against the side of the cockpit; a sharp pain radiated through her arm. She finished the assignments and selected ENGAGE. Lock-lines leapt from the Bots’ icons, attaching themselves to the track-files. She looked out the canopy; she could see the bots now—small dark dots approaching from the blue. They grew bigger, racing towards the target at maximum speed. Small bombs began to fall away from them.
On her map display, the lock-lines flashed, indicating weapon release. ‘60 seconds to impact!’ Jazz yelled. She pulled up one of the Bot sensor feeds; red boxes outlined each enemy soldier. They were running, flitting between the trees. Maybe they heard the scream of the approaching Bots. She held her breath, ‘15 seconds.’
BOOM! BO-BOOM, BOOM! Explosions appeared, shock waves ripping through the trees, their intensity gaining out the picture. She looked out below her; dark smoke was rising from 10 impact points scattered across the terrain. Glancing up, she caught sight of the Bots streaking off into the distance, disappearing back into the blue.
They spent a few minutes watching the Bot sensor feeds, looking for survivors. ‘Not detecting any spitters,’ Vector said, holding a thumbs up against the canopy for her to see. ‘Nice job!’
‘Cannon 1, this is Havoc. The patrol reports good hits. Engagement complete,’ said the JTAC confirming their assessment. In the background, another voice at the JTAC’s headquarters started talking urgently. He let go of the transmit switch, and his comms cut out.
They continued to circle, letting the Bots settle back into their overwatch scan. After a few minutes, the radio crackled. ‘Cannon, this is Havoc. We have indications of enemy reinforcements gathering at a bridge to the south. I’m sending you the coordinates.’
A point-of-interest icon appeared on Jazz’s display; it was a bridge seven kilometres south of the patrol, just on the outskirts of a small town. ‘I’ll reassign the Bots,’ she advised Vector, clicking through menu options. She watched as their icons repositioned to the south.
‘I don’t think we’ll get a camera in there, Jazz,’ Vector said. ‘The gaps in the cloud are closing up.’ Jazz looked out the canopy. He was right; the low cloud was getting a lot thicker.
‘Rog, you take Bot 4. See if you can get it under,’ she replied. She zoomed in around the bridge on her display. Soon, a yellow track-file appeared on it, then four more aligned in a column facing north. Cursoring over them, she saw the contributor was LW-Radar - Prob Technical (76%), indicating a lower confidence radar classification by the Bots.
The JTAC’s voice returned: ‘Cannon, intelligence indicates we’ve got six technicals on the bridge. We expect they are preparing a second attack on the patrol’s position. Can you get eyes on to confirm the target and assess collateral?’
‘Negative,’ Vector said to Jazz. ‘I’ve tried. The cloud is too low; we can’t get a sensor under it.’ Jazz passed that onto the JTAC.
‘Roger,’ the JTAC replied. ‘We’ll see what we can do; we need to engage before they move north and separate.’
Jazz moved her cursor from track to track. One of them switched to red, she snapped to it—MANDELA. The lead track-file began to move. Those still on the bridge turned red, MANDELA recommending their engagement. ‘Havoc, this is Cannon,’ Jazz said quickly, ‘One of the tracks has started moving north.’
‘Havoc, copied,’ the JTAC replied. ‘Stay with the other vehicles. We’re discussing it here.’
Jazz and Vector continued to orbit over the target area for several minutes in silence. Surely they’ll engage, Jazz thought; MANDELA has information we don’t have. She leant her helmet against her seat’s headrest, frustrated at the delay and feeling powerless to prevent the impending attack on the patrol—she wouldn’t release without JTAC approval.
The radio sparked to life; she sat bolt upright. ‘Cannon, this is Havoc; you are cleared to engage MANDELA tracks, ROE 556, Ground Force Commander initials Mike Papa.’
‘Thank god! Expect Wingmen 3 and 4 again,’ Jazz told Vector. ‘Airburst, radar guidance, auto. Your spitters!’ Jazz quickly set the target assignments, lock-lines snapping from the Bots to the red tracks on the bridge. She turned her attention from the screen to the real world outside, looking for their approach. The Bots appeared again out of the blue, one from the north and one from the south—dark, angry dots racing towards the target. ‘60 seconds,’ she called.
Now only a few miles away, the Bots streaked towards the target head-head with each other, their bombs falling towards the cloud below them. They both broke away in a hard right turn, passing left-to-left, only avoiding collision by a few hundred metres. Jazz’s eyes followed the bombs as they fell into the cloud. She keyed the radio, ‘15 seconds.’
Four years later, 09:37 07 Nov 2044, 60km/h southbound on the Kings Ave Bridge
‘Oh my god! Have you read the news today?!’ Jazz’s voice blared through Charles’ audio implant.[v] Charles opened a projection in the sedan’s back seat and started flicking through the news stories. His autonomous car sped over the King’s Avenue bridge towards Parliament House; pedestrians enjoying the clear summer’s day flashed past the window. He was the Deputy Chief of Staff to CDF, en route to help the boss through the Senate inquiry. Jazz was his close friend; they’d known each other since attending ADFA 21 years ago.
‘I’ve got it, I’ve got it. Reading it now,’ he replied.
The Sydney Morning Post
War in the Pacific: Australian Top Guns kill 93 in bungled mission
Tuesday, November 7th, 2044
By Laura Timms
An Australian F/A-18F crew stands accused of killing 93 innocent refugees and wounding 106 others in a bungled AI-led mission in the Philippines four years ago.
Three months ago, the 2040 incident came to light when a group of survivors began contesting the Philippine government’s narrative that the casualties were caused by a separatist IED. They claimed that jet noise could be heard before the explosion.
Liberté, the French human rights group, analysed mass data to piece together the incident, proving an Australian F/A-18F aircraft, callsign Cannon 1, and its four Loyal Wingmen drones were involved in the killings.
“…There’s no way on that day they couldn’t have seen the… civilians.”
The Australian Chief of Defence Force (CDF) has stated that the strike’s intended target was 6 separatist technicals. He said the ADF’s own investigation is ongoing. The CDF’s purported narrative clashes with the traumatic account of incident survivors. Amani Patino was standing near the Puerto Princesa highway bridge outside the town of Sofronio Española. He told SMP that hundreds of families, women and children were trying to cross. They were blocked by a single separatist standing in the back of a truck firing his rifle into the air.
Dr Mahmoud Assaf, an aid worker for Islamic Relief, responded to the strike: “At the time, they told us an IED had exploded. When we arrived I saw women and children screaming, it was chaos... I was working on one of the wounded children; there was nothing we could do for her, she kept breathing for hours. It was torture for her parents because they knew she wouldn’t live… It’s tragic to hear it was caused by Australian aircrew. There’s no way on that day they couldn’t have seen the large crowd of civilians.”
SMP investigators have spoken with an ADF whistle-blower who passed SMP a classified “Battle Damage Report.” It stated a “militant convoy” was positively identified by the ADF’s MANDELA AI, a controversial intelligence fusion and decision-making assistant acquired from the US government. The report said weather prevented the F/A-18F aircrew from seeing the bridge themselves. That raises the question - why were the aircrew blindly trusting a computer’s recommendation to drop weapons into a populated town?
“I was working on one of the wounded children; there was nothing we could do for her, she kept breathing for hours…”
Senator Susan Porter from the Greens Party said the aircrew must be held to account: “Enough is enough; ADF officers must be held responsible for their actions. A criminal investigation must begin immediately. It’s not good enough for them to blame their errors on a computer.”
SMP is continuing to investigate. Who – or what – confirmed the information that triggered the strike? How could the F/A-18F aircrew justify dropping their bombs on a target they hadn’t seen?
More to follow.
The Morning Courier
Nick Meister
2 min read
November 7, 2044 - 8:32AM
Opinion: ADF’s AI Command Vacuum
Key Points: It is alleged that a MANDELA AI-led RAAF air strike on 04 Feb 2040 killed 93 and wounded 106 civilian refugees. MANDELA is an artificial intelligence system designed to support commanders’ decision making. It was purchased from the US in 2036. A Senate inquiry into the AI killings will begin today. |
The success of the ADF’s combat operations in the Philippines has been overshadowed by revelations of devastating official incompetence and callousness that has called hundreds of airstrikes into question. It has come to light that the ADF regularly accepted targeting recommendations from their MANDELA AI, the cause of the bridge attack that killed 93 civilians in Sofronio Española. Members on all sides of politics are calling for an immediate, third-party inquiry into what could amount to widespread war crimes.
A leaked “Fact-Finding Report” gave details on the highly classified Machine Analysis Decision Learning Algorithm (MANDELA). In it, the US-manufactured AI is described as “a highly complex system that analyses information from sensors, military intelligence sources and open sources. The system is designed to assess thousands of pieces of information, from mobile phone movements to social media collections to personnel movements through ambient electromagnetic radiation. It allows military targets to be identified and engaged in highly complex environments well within human decision timelines.”
The ADF purchased the MANDELA AI from the US government in 2036 to support command decision making in combat and aid US interoperability. The AI was used to recommend hundreds of coalition strikes in the Philippines. Military commentators argued its employment was essential for targeting an enemy largely indistinguishable from the civilian population.
However, using a decision AI system is not without risk. US political-military strategist, Dr Thomas K Adams of John Hopkins University, highlighted that the involvement of an AI “creates an environment too complex for humans to direct… producing a data overload that makes it impossible for humans to directly intervene in decision-making.”[vi] That risk appears to have been tragically realised during the Feb 2040 strike – and possibly hundreds more.
The leaked report indicates that the ADF had little understanding of the inputs used by MANDELA, the accuracy of those inputs or the operation of its analytics.[vii] This raises the question, why was the ADF using a system they had such little knowledge of?
“…too complex for humans to direct… producing a data overload that makes it impossible for humans to directly intervene in decision making…”
Clearly, this shows a lack of accountability by the ADF. It appears little effort was spent on establishing safeguards, policy or procedures for the system. Who was responsible for overseeing the MANDELA AI’s decision making? Who is accountable for the deaths of these civilians? Today’s Senate inquiry will seek to answer those questions.
‘Oh, Jazz. I’m so sorry...’ Charles said.
‘The news feed is full of them!’ she interrupted, her voice wavering. ‘They’re now calling Vector and I murderers—you know that?! We followed all the procedures—we wouldn’t have dropped if we knew there were civilians on that bridge!’
‘I know, Jazz, you don’t need to convince me,’ Charles replied. There was silence on the other end of the call; he could tell she had tears in her eyes. A few months ago, the incident had hit the media, a surprise to the ADF, which thought its Philippine operations were clean. Those involved had been stood down while Defence conducted its investigation; Jazz had lost her squadron command. Today the news commentary had intensified, accusing the aircrew of being war criminals.
Charles looked out the windscreen as the car climbed Capital Hill. ‘We’ll sort this out, Jazz,’ he assured her. ‘Let’s see what comes out of this Senate inquiry; command knows you didn’t do anything wrong—they have your back.’
‘I don’t know what I’m going to do,’ she said. ‘Even if it does get cleared up, my life is over. Do know you Lucy got harassed at school yesterday? She came home crying, saying “Mummy murdered people!”’
‘Hey, hey, hey! We’ll sort it out. I’m just pulling up to Parliament now. I’ll let you know how it goes. You hang in there, alright?’ Jazz murmured an acknowledgement and hung up. Charles worried about her; she was a good person. He couldn’t imagine what she was going through.
He hurried through the corridors towards the Senate, squeezing past people enjoying their morning coffees. He clutched his classified tablet tightly; it contained the MANDELA incident information in case CDF needed to refer to it.
For the last two weeks, Charles had gone through it all: the after-action reports, the witness statements, the military investigation into MANDELA, the project files—it was a mess. There was little detailed technical understanding of how the system worked and thousands of scenario variations that weren’t tested.[viii] With the benefit of hindsight, anyone could see this incident was bound to happen eventually—but what other option did the ADF have in the AI era? The advantage MANDELA provided to targeting was immense. If the system outperformed human analysts by a wide margin, was that acceptable, even if it wasn’t perfect? It was evident the media—and certain senators—thought not.
Charles pushed open the heavy wooden door; the room smelled of polished wood and leather. Staff bustled about preparing microphones before the arrival of the senators, CDF and the Secretary. The senators’ seats were at the front of the room: two rows of high-backed leather chairs behind large wooden desks, elevated like a grandstand. They commanded the room, their authority cemented by the Australian flags behind them. In front of their seats were two small desks for CDF and the Secretary. Charles shuddered, wondering if they’d tried to find cheaper chairs to further emphasise the theatre of subservience. He wouldn’t like to be in CDF’s position. He quickly strode to the seats arranged behind CDF’s desk and sat down next to the Secretary’s staffers, two clean-cut men in dark suits. They gave Charles anxious nods but quickly returned to fixate on their tablets.
Charles fidgeted with his tie and checked the time; 10 minutes to go, CDF and the Secretary would be arriving soon. He ensured his tablet was working and reran the information in his mind. His team had tried to anticipate each senator’s questions given their political agendas. The Chair, Senator Mathew Sainsbury, a Liberal, would likely be supportive; his government had approved MANDELA in 2036. Senator Susan Porter, Greens, would probably try to frame the CDF as a militaristic warmonger, with much emotion but little rationale. Senator Jamie Martin would be their biggest problem. She was a Labor senator with a background in mathematics and software engineering; she would use this opportunity to prove the Liberal ministers were incompetent at supervising their departments.
The door burst open, a stream of people moved in, and the air filled with chatter. Charles watched as CDF and the Secretary walked in together. Striding towards him, CDF met his eyes. ‘You all good?’ he inquired, eyebrow raised, placing his documents on the desk. Surprisingly, the CDF seemed calm and collected. Charles gave him a nervous nod.
The senators filed in, taking their places. Charles’ mind drifted to the phone call with Jazz. He knew she’d be watching the session live. He was worried about how she would take it.
The chatter subsided as the last Senator sat down. The Chair, the elderly Senator Sainsbury, leaned close to the microphone and adjusted his glasses. He made a few opening comments, followed by several perfunctory questions to establish an overview of the incident. He opened the floor for the other senators to ask questions.
Senator Porter moved to speak first. She pushed her long dark hair behind her shoulders. ‘CDF,’ she said, frowning, ‘this is the worst war crime in recent history. How did this AI, and by implication the ADF, fail so badly?’
The CDF calmly leaned closer to his microphone: ‘Senator, MANDELA uses a large collection of incidental data sources to produce its determinations: mobile phone data, wifi analysis, internet activity, for example. In this case, many of those sources were not available.[ix] The separatists had destroyed the phones in that village and cut off its connectivity to the outside world. Further complicating matters, we believe a foreign actor was spoofing civilian transmissions in the area…’.[x]
‘And what?’ Senator Porter interrupted aggressively: ‘You knew this and trusted the system’s recommendations anyway?’
‘No, Senator,’ CDF replied, ‘we didn’t know at the time. We only learnt that after Northwell Atomics, the prime contractor for MANDELA, completed their investigation and analysis this year. They informed us that due to foreign spoofing MANDELA failed to recognise that civilians in that area could no longer be tracked via electronic emissions. Over several days its learning algorithms had classified the spoofing as a benign part of the ambient environment. Each successful recommendation made in the spoofing’s presence encouraged it to reinforce that assumption.’
‘What oversight did you have of these learning algorithms?’ Porter replied. ‘Surely you could have detected this issue.’
‘Senator, we had protocols to impose safeguards,’ said CDF. ‘However, we had no meaningful visibility into MANDELA’s decision processes. They are the proprietary intellectual property of Northwell Atomics. They are also covered by the United States’ International Traffic of Arms Regulations, which restricts the release of that information to foreigners.’[xi]
Senator Porter leaned forward menacingly: ‘You killed 93 civilians, General,’ she said in a low, dark tone, slowly increasing in volume. ‘Why on earth do you think it’s acceptable to use an AI into which you have no visibility or understanding?’
‘Senator,’ CDF started, raising his palms disarmingly.
‘93!’ she shouted, her face now reddening. ‘Men, women, children! It’s a bloody war crime, General!’ She slammed her fist on the desk. The CDF paused, waiting for Senator Porter to calm down.
‘Senator,’ CDF restarted, ‘we recognise the gravity of this tragedy too, but the limitations of our knowledge were understood and accepted. All ADF weapons proceed through a detailed capability acquisition system that identifies the risks and benefits for the consideration of the National Security Committee. The benefits of MANDELA were assessed to outweigh the risks.’
He raised his hands to explain. ‘You see, human decision-makers fail as well; they are subject to biases and error, and these have also led to disaster. In contrast, MANDELA has been statistically proven to outperform its human counterparts; it identifies and engages targets with greater speed and accuracy. In high-end combat scenarios, MANDELA is vital. In low-level contexts, like the Philippines, MANDELA saves lives. It has saved civilian lives by accurately analysing collateral damage and saved military lives by permitting target engagement before they endanger friendly forces.’ He glanced down at his notes.
‘Even just moments before the incident,’ he continued, ‘the same F/A-18F aircrew had successfully engaged and killed 32 enemy combatants who were advancing on an AFP patrol. Without MANDELA, those enemy combatants would have attacked the patrol, and friendly lives would have been lost.’ He paused, letting that sink in.
Senator Porter shifted uncomfortably, frustrated. She didn’t have the technical knowledge to contest CDF’s assertions—and she knew it. Senator Jamie Martin, however, sitting at the opposite end of the back row, looked decidedly unimpressed. She sat forward and calmly turned over a page in front of her. ‘So, General,’ she started, ‘what was the justification for acquiring this system back in 2036?’
‘At the time, China had just put its own AI system into service, Māotóuyīng,’[xii] the CDF replied. ‘To ensure its allies remained capable and competitive, the US Congress approved the release of MANDELA to Australia, Japan, and South Korea. Our modelling showed that without an AI, we would be combat ineffective against the Chinese; humans simply can’t make decisions fast enough in modern warfare.[xiii] The information is too complex, and the engagement timelines are too short.’ Senator Martin listened intently, nodding her head in agreement.
The CDF continued: ‘The US had also made it clear that MANDELA was vital for continued US-Australian interoperability by providing an interface to the American MANDELA system.[xiv] This would allow Australia to contribute to coalition warfighting effectively and would reduce the possibility of American and Australian fratricide. In exchange, we would have access to all US intelligence feeds connected to their system. We explored alternatives—there were none. If we wanted to remain militarily relevant in the region, we had no choice—the NSC understood that.’
‘Yes, I can see its benefits there,’ Senator Martin replied, thoughtfully nodding, ‘but yet… the Philippines was hardly a high-end conflict. It was a low-tech separatist movement, wasn’t it?’
‘Yes, Senator. Therefore, we didn’t have MANDELA set to conduct automatic engagements; it just presented recommendations to the commanders for their decision. So, we managed the risk there.’
Senator Martin appeared bemused; the corner of her mouth tilted up in a wry smile. ‘You managed the risk there…’, she echoed to herself softly, looking down and rubbing the desk in a small circular motion. Her hand stopped, and she looked directly at him. ‘Tell me: before this point, how often had these commanders seen MANDELA fail?’
‘Never,’ CDF replied proudly.
She smiled. ‘So, when presented with a MANDELA targeting recommendation against an imminent threat, under what conditions would your commanders reject it?’
The CDF looked down at his notes, considering his response deeply. She continued before he could answer: ‘Would you say our detailed capability acquisition system has reduced them to servants of a computer they have no understanding of?’[xv]
Four weeks later, 11:58 05 Dec 2044, 0kts by Lake Burley-Griffin
Charles stood by the coffee cart, looking out over the lake. It was a bright blue day, the sun reflecting off the water into his eyes. The brewing coffee smelt good. He turned towards the car park behind him. Jazz, her husband Mike and their ten year old Lucy were just getting out of their car. Charles gave them a wave. His own wife and son were already up at the playground.
Mike waved, and Jazz caught sight of him. She walked down towards him. ‘Let me help you with the coffees,’ she said with a slight smile. ‘Mike will take Lucy up.’
A group of lunchtime joggers ran past. ‘Are you doing alright?’ Charles asked.
Jazz gave a forced smile. ‘Yeah, alright,’ she replied quietly as they joined the coffee line. ‘I’m still posted to Canberra next year—a job in VCDF’s Complex Systems Review Team. They’re trying to work with the US government and industry to gain more oversight of their combat systems.’
‘Nice. Well, I imagine it’s good to be getting back to work, and it’ll be great to make a difference.’
Jazz shook her head slightly, looking down at her feet. They stood in silence for a while, listening to the coffee cart grinding beans. She gently turned her foot in the loose gravel and looked up at him. ‘It’s a little too close to home… you know?’
Endnotes
[i]Malcolm Davis, ‘The Role of Autonomous Systems in Australia’s Defence,’ After Covid-19: Australia and the world rebuild (Volume 1) (2020): 106-09.
[ii]Malcolm Davis, ‘Loyal Wingman Leads the Way to the RAAF of 2121,’ The Strategist, Australian Strategic Policy Institute, 2021.
[iii]Johan Schubert et al., ‘Artificial Intelligence for Decision Support in Command and Control Systems’ (paper presented at the 23rd International Command and Control Research & Technology Symposium Multi-Domain C, 2018), 6-9.
[iv]Marta Bistron and Zbigniew Piotrowski, ‘Artificial Intelligence Applications in Military Systems and Their Influence on Sense of Security of Citizens,’ Electronics (Basel) 10, no. 871 (2021): 11-12, https://doi.org/10.3390/electronics10070871.
[v]Natalie Moreton, ‘What Does the Future Hold for the Hearables Market? IDTechEx Reveals,’ PR Newswire (New York), 11/02/2020 2020, 1-3.
[vi]Thomas K Adams, ‘Future Warfare and the Decline of Human Decision making,’ The US Army War College Quarterly: Parameters 41, no. 4 (2011): p. 2.
[vii]Amina Adadi and Mohammed Berrada, ‘Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI),’ IEEE access 6 (2018): 1-2.
[viii]Greg Allen, nderstanding AI technology, Joint Artificial Intelligence Center (JAIC) The Pentagon United States (2020), 4.
[ix]Artificial Intelligence and National Security, Federation of American Scientists (2020), 16-17, ttps://go.exlibris.link/px4xDRkr.
[x]M. A. Thomas, ‘Time for a Counter-AI Strategy,’ Strategic Studies Quarterly 14, no. 1 (2020): 6, https://www.jstor.org/stable/26891880.
[xi]C. Thomsen Roszel, II, ‘Artificial Intelligence and Export Controls: Conceivable, but Counter productive?,’ 22, no. 5 (2018): 14-15.
[xii]Gregory C. Allen, Understanding China's AI Strategy: Clues to Chinese Strategic Thinking on Artificial Intelligence and National Security, Center for a New American Security (2019), 3-4.
[xiii]Glenn Moy et al., Recent Advances in Artificial Intelligence and their Impact on Defence, 25 (Canberra 2020).
[xiv]Simona R. Soare, Digital divide? Transatlantic Defence Cooperation on Artificial Intelligence, European Union Institute for Security Studies (EUISS) (2020), 5-8, http://www.jstor.org/stable/resrep25027.
[xv]Thomas, ‘Time for a Counter-AI Strategy,’ 4.
Comments
Start the conversation by sharing your thoughts! Please login to comment. If you don't yet have an account registration is quick and easy.