App-based companies have publicly spoken for years about the money-saving potential of autonomous vehicles. These firms have poured billions of dollars into recruiting and managing the independent contractors who do the delivering and driving for them, and millions more ensuring they’ll stay independent contractors and not employees. What if the firms could skip all that? What if robots did all the work, or at least some of it?
Still, with today’s announcement, DoorDash is throwing its lot into an industry that has faced some choppy waters. And, of course, the threat of public kicks.
Speed Bumps
Delivery robots were hyped during the onset of the Covid pandemic as a solution to that other very human problem of contagion. Since then, however, Amazon and FedEx abandoned their delivery robot projects; others working on delivery bots have pivoted to software or industrial uses. The companies that remain have mostly focused on smaller deployments on college campuses or a select few cities, and those don’t seem to be growing as quickly as hoped.
Estonian company Starship Technologies, the biggest one still standing in the delivery robot space, has found a niche operating on mostly university campuses, where streets and sidewalks are wide, well-maintained, and relatively friendly, and where those desperately seeking 2 am pizzas and burritos are at their least price-sensitive. Postmates spinoff Serve Robotics launched in 2017 but has built only 400 robots, according to its most recent financial flings, with goals to build 2,000 by the end of the year.
Contrast that with the growth in autonomous vehicles. Though robotaxi services are still limited to a handful of global cities, they’re picking up and dropping off customers to the tune of hundreds of thousands of rides per week.
The reason for the slower growth in delivery robots is actually pretty simple, says Bern Grush, the executive director of the nonprofit Urban Robotics Foundation: “You’re trying to solve a much harder problem with far, far, far less capital and far, far, far less compute.”
Consider the technical challenge Dot has ahead of it: DoorDash says the robot is built to operate on sidewalks, bike lanes, and roads. It’s meant to pilot in and out of parking lots to pick up food and to navigate driveways and apartment complexes to drop it off. That means the software needs to “understand,” predict the movement of, and get around a remarkable number of situations, vehicles, and living things: cars, trucks, school buses, strollers, children’s bicycles, aggressive mopeds, motorized wheelchairs, dogs, squirrels, toddlers, people with limps, runners, competitive cyclists. And on and on.
Each Dot can carry 30 pounds of cargo, drive up to 20 mph, and go about five miles on a charge.
A study by Boston Consulting Group (BCG) suggests that organisations that lead in technology development are gaining a first-mover advantage when it comes to artificial intelligence (AI) and using agentic AI to improve business processes.
What is striking about BCG’s findings, according to Jessica Apotheker, managing director and senior partner at Boston Consulting Group, is that the leading companies in AI are mostly the same ones that were leaders eight years ago.
“What this year’s report shows is that the value gap between these companies and others is widening quite a bit,” she says. In other words, BCG’s research shows that organisations that have invested disproportionately in technology achieve a higher return from that investment.
Numerous pieces of research show that a high proportion of AI initiatives are failing to deliver measurable business success. BCG’s Build for the future 2025 report shows that the companies it rates as the best users of AI generate 1.7 times more revenue growth than the 60% of companies in the categories it defines as stagnating or emerging.
However, there is certainly a case to build on previous initiatives such as robotic process automation (RPA).
Speaking at the recent Forrester Technology and Innovation Summit in London, Bernhard Schaffrik, principal analyst at Forrester, discussed how agentic AI can be built on top of a deterministic RPA system to provide greater flexibility than what existing systems can be programmed to achieve.
The analyst firm uses the term “process orchestration” to describe the next level of automating business processes, using agentic AI in workflow to handle ambiguities far more easily than the programming scripts used in RPA.
“Classic process automation tools require you to know everything at the design stage – you need to anticipate all of the errors and all the exceptions,” says Schaffrik.
He points out that considering these things at design time is unrealistic when trying to orchestrate complex processes. But new tools are being developed for process orchestration that rely on AI agents.
A strong data foundation
Boston Consulting Group (BCG) says prerequisites for the successful roll-out of AI agents include strong data foundations, scaled AI capabilities and clear governance.
Standardisation of data is a key requirement for success, according to Twig. “A big part of the issue is data,” he says. “AI is only as strong as the information it runs on, and many companies don’t have the standardised, consistent datasets needed to train or deploy it reliably.”
Within the context of agentic AI, this is important to avoid miscommunications both at the technology infrastructure level and in people’s understanding of the information. But the entire data foundation does not have to be built all at once.
BCG’s Apotheker says companies can have an enterprise-wide goal to achieve clean data, and build this out one project at a time, providing a clean data foundation on which subsequent projects can be built. In doing so, organisations are able to gain a better understanding of the enterprise data these projects require while they ensure that the datasets are clean and good data management practices are followed.
A working agentic AI strategy relies on AI agents connected by a metadata layer, whereby people understand where and when to delegate certain decisions to the AI or pass work to external contractors. It’s a focus on defining the role of the AI and where people involved in the workflow need to contribute.
This functionality can be considered a sort of platform. Scott Willson, head of product marketing at xtype, describes AI workflow platforms as orchestration engines, coordinating multiple AI agents, data sources and human touchpoints through sophisticated non-deterministic workflows. At the code level, these platforms may implement event-driven architectures using message queues to handle asynchronous processing and ensure fault tolerance.
Data lineage tracking should happen at the code level through metadata propagation systems that tag every data transformation, model inference and decision point with unique identifiers. Willson says this creates an immutable audit trail that regulatory frameworks increasingly demand. According to Willson, advanced implementations may use blockchain-like append-only logs to ensure governance data cannot be retroactively modified.
Adapting workflows and change management
Having built AI-native systems from the ground up and transformed the company’s own product development processes using AI, Alan LeFort, CEO and co-founder of StrongestLayer, notes that most organisations are asking completely the wrong questions when evaluating AI workflow platforms.
“The fundamental issue isn’t technological, it’s actually organisational,” he says.
Conway’s Law states that organisations design systems that mirror their communication structures. But, according to LeFort, most AI workflow evaluations assume organisations bolt AI onto existing processes designed around human limitations. This, he says, results in serial decision-making, risk-averse approval chains and domain-specific silos.
When you try to integrate AI into human-designed processes, you get marginal improvements. When you redesign processes around AI capabilities, you get exponential gains Alan LeFort, StrongestLayer
“AI doesn’t have those limitations. AI can parallelise activities that humans must do serially, doesn’t suffer from territorial knowledge hoarding and doesn’t need the elaborate safety nets we’ve built around human fallibility,” he adds. “When you try to integrate AI into human-designed processes, you get marginal improvements. When you redesign processes around AI capabilities, you get exponential gains.”
StrongestLayer recently transformed its front-end software development process using this principle. Traditional product development flows serially. A product manager talks to customers, extracts requirements and then hands over to the user experience team for design, the programme management team then approves the design, and developers implement the software. It used to take 18-24 months to completely rebuild the application in this process, he says.
Instead of bolting AI onto this process, LeFort says StrongestLayer “fundamentally reimagined it”.
“We created a full-stack prototyper role-paired with a front-end engineer focused on architecture. The key was building an AI pipeline that captured the contextual knowledge of each role: design philosophy, tech stack preferences, non-functional requirements, testing standards and documentation needs.”
As a result of making these workload changes, he says the company was able to achieve the same outcome from a product development perspective in a quarter of the time. This, he says, was not necessarily achieved by working faster, but by redesigning the workflow around AI’s ability to parallelise human sequential activities.
LeFort expected to face pushback. “My response was to lead from the front. I paired directly with our chief product officer, Joshua Bass, to build the process, proving it worked before asking others to adopt it. We reframed success for our front-end engineer around velocity and pioneering new ways of working,” he says.
For LeFort, true speed to value comes from two fundamental sources: eliminating slack time between value activities and accelerating individual activity completion through AI automation. “This requires upfront investment in process redesign rather than quick technology deployment,” he says.
LeFort urges organisations to evaluate AI workflow platforms based on their ability to enable fundamental process transformation, rather than working to integrate existing inefficiencies.
Getting agentic AI decision-making right
Research from BCG suggests that the best way to deploy agents is through a few high-value workflows with clear implementation plans and workforce training, rather than in a massive roll-out of agents everywhere at once.
There are different models with different strengths. We want to use the best model for each task Ranil Boteju, Lloyds Banking Group
One of the areas IT leaders need to consider is that their organisation will more than likely rely on a number of AI models to support agentic AI workflows. For instance, Ranil Boteju, chief data and analytics officer at Lloyds Banking Group, believes different models can be tasked with tackling each distinct part of a customer query.
“The way we think about this is that there are different models with different strengths, and what we want to do is to use the best model for each task,” says Boteju. This approach is how the bank sees agentic AI being deployed.
With agentic AI, problems can be broken down into smaller and smaller parts, where different agents respond to each part. Boteju believes in using AI agents to check the output from other agents, rather like acting as a judge or a second-line colleague acting as an observer. This can help to cut erroneous decision-making arising from AI hallucinations when the AI model basically produces a spurious result.
IT security in agentic AI
People in IT tend to appreciate the importance of adhering to cyber security best practices. But as Fraser Dear, head of AI and innovation at BCN, points out, most users do not think like a software developer who keeps governance in mind when creating their own agents. He urges organisations to impose policies that ensure the key security steps are not skipped in the rush to deploy agentic AI.
“Think about what these AI agents might access across SharePoint: multiple versions of documents, transcripts, HR files, salary data, and lots more. Without guardrails, AI agents can access all this indiscriminately. They won’t necessarily know which versions of these documents are draft and which are approved,” he warns.
The issue escalates when an agent created by one person is made available to a wider group of colleagues. It can inadvertently give them access to data that is beyond their permission level.
Dear believes data governance needs to include configuring data boundaries, restricting who can access what data according to job role and sensitivity level. The governance framework should also specify which data resources the AI agent can pull from.
In addition, he says AI agents should be built for a purpose, using principles of least privilege: “Just like any other business-critical application, it needs to be adequately tested and ‘red-teamed’. Perform penetration testing to identify what data the agent can surface, to whom, and how accurate the data is. Track and audit which agents are accessing which data and for what purpose, and implement real-time alerts to flag unusual access patterns.”
A bumpy ride ahead
What these conversations with technology experts illustrate is that there is no straightforward path to achieving a measurable business benefit from agentic AI workflows – and what’s more, these systems need to be secure by design.
Organisations need to have the right data strategy in place, and they should already be well ahead on their path to full digitisation, where automation through RPA is being used to connect many disparate workflows. Agentic AI is the next stage of this automation, where an AI is tasked with making decisions in a way that would have previously been too clunky using RPA.
However, automation of workflows and business processes are just pieces of an overall jigsaw. There is a growing realisation that the conversation in the boardroom needs to move beyond the people and processes.
BCG’s Apotheker believes business leaders should reassess what is important to their organisation and what they want to focus on going forward. This goes beyond the build versus buy debate: some processes and tasks should be owned by the business; some may be outsourced to a provider that may well use AI; and some will be automated through agentic AI workflows internally.
It is rather like business process engineering, where elements powered by AI sit alongside tasks outsourced to an external service provider. For Apotheker, this means businesses need to have a firm grasp of what part of the business process is strategically important and can be transformed internally.
Business leaders then need to figure out how to connect the strategically important part of the workflow to what the business actually outsources or potentially automates in-house.
AI’s superior ability to formulate thoughts and statements for us weakens our judgment and ability to think critically, says media professor Petter Bae Brandtzæg.
No one knew about Chat GPT just three years ago. Today, 800 million people use the technology. The speed at which AI is rolling out breaks all records and has become the new normal.
Many AI researchers, like Brandtzæg, are skeptical. AI is a technology that interferes with our ability to think, read, and write. “We can largely avoid social media, but not AI. It is integrated into social media, Word, online newspapers, email programs, and the like. We all become partners with AI—whether we want to or not,” says Brandtzæg.
The professor of media innovations at the University of Oslo has examined how AI affects us in the recently completed project “An AI-Powered Society.”
The freedom of expression commission overlooked AI
The project has been conducted in collaboration with the research institute SINTEF. It is the first of its kind in Norway to research generative AI, that is, AI that creates content, and how it affects both users and the public.
The background was that Brandtzæg reacted to the fact that the report from the Norwegian Commission for Freedom of Expression, which was presented in 2022, did not sufficiently address the impact of AI on society—at least not generative AI.
“There are studies that show that AI can weaken critical thinking. It affects our language, how we think, understand the world, and our moral judgment,” says Brandtzæg.
A few months after the Commission for Freedom of Expression report, ChatGPT was launched, making his research even more relevant.
“We wanted to understand how such generative AI affects society, and especially how AI changes social structures and relationships.”
AI-Individualism
The social implications of generative AI is a relatively new field that still lacks theory and concepts, and the researchers have therefore launched the concept of “AI-individualism.” It builds on “network individualism,” a framework which was launched in the early 2000s.
Back then, the need was to express how smartphones, the Internet, and social media enabled people to create and tailor their social networks beyond family, friends, and neighbors.
Networked individualism showed how technology weakened the old limits of time and place, enabling flexible, personalized networks. With AI, something new happens: the line between people and systems also starts to blur, as AI begins to take on roles that used to belong to humans.
“AI can also meet personal, social, and emotional needs,” says Brandtzæg.
With a background in psychology, he has for a long time studied human-AI relationships with chatbots like Replika. ChatGPT and similar social AIs can provide immediate, personal support for any number of things.
“It strengthens individualism by enabling more autonomous behavior and reducing our dependence on people around us. While it can enhance personal autonomy, it may also weaken community ties. A shift toward AI-individualism could therefore reshape core social structures.”
He argues that the concept of “AI-individualism” offers a new perspective for understanding and explaining how relationships change in society with AI. “We use it as a relational partner, a collaborative partner at work, to make decisions,” says Brandtzæg.
Students choose chatbot
The project is based on several investigations, including a questionnaire with open-ended answers to 166 high school students on how they use AI.
“They (ChatGPT and MyAI) go straight to the point regarding what we ask, so we don’t have to search endlessly in the books or online,” said one high school student about the benefits of AI.
“ChatGPT helps me with problems, I can open up and talk about difficult things, get comfort and good advice,” responded a student.
In another study, using an online experiment with a blind test, it turned out that many preferred answers from a chatbot over a professional when they had questions about mental health. More than half preferred answers from a chatbot, less than 20% said a professional, while 30% responded both.
“This shows how powerful this technology is, and that we sometimes prefer AI-generated content over human-generated,” says Brandtzæg.
‘Model power’
The theory of “model power” is another concept they’ve launched. It builds on a power relationship theory developed by sociologist Stein Bråten 50 years ago.
Model power is the influence one has by being in possession of a model of reality that has impact, and which others must accept in the absence of equivalent models of power of their own, according to the article “Modellmakt og styring” (online newspaper Panorama—in Norwegian).
In the 1970s, it was about how media, science, and various groups with authority could influence people, and had model power. Now it’s AI.
Brandtzæg’s point is that AI-generated content no longer operates in a vacuum. It spreads everywhere, in public reports, new media, in research, and in encyclopedias. When we perform Google searches, we first get an AI-generated summary.
“A kind of AI layer is covering everything. We suggest that the model power of social AI can lead to model monopolies, significantly affecting human beliefs and behavior.”
Because AI models, like ChatGPT, are based on dialog, they call them social AI. But how genuine is a dialog with a machine fed with enormous amounts of text?
“Social AI can promote an illusion of real conversation and independence—a pseudo-autonomy through pseudo-dialog,” says Brandtzæg.
Critical but still following AI advice
According to a survey from The Norwegian Communications Authority (Nkom) from August 2025, 91% of Norwegians are concerned about the spread of false information from AI services like Copilot, ChatGPT, and Gemini.
AI can hallucinate. A known example is a report the municipality of Tromsø used as a basis for a proposal to close eight schools, was based on sources that AI had fabricated. Thus, AI may contribute to misinformation, and may undermine user trust in both AI, service providers and public institutions.
Brandtzæg asks how many other smaller municipalities and public institutions have done the same and he is worried about the spread of this unintentional spread of misinformation.
He and his researcher colleagues have reviewed various studies indicating that although we like to say we are critical, we nevertheless follow AI’s advice, which highlights the model power in such AI systems.
“It’s perhaps not surprising that we follow the advice that we get. It’s the first time in history that we’re talking to a kind of almighty entity that has read so much. But it gives a model power that is scary. We believe we are in a dialog, that it’s cooperation, but it’s one-way communication.”
American monoculture
Another aspect of this model power is that the AI companies are based in the U.S. and built on vast amounts of American data.
“We estimate that as little as 0.1% is Norwegian in AI models like ChatGPT. This means that it is American information we relate to, which can affect our values, norms and decisions.”
What does this mean for diversity? The principle is that “the winner takes it all.” AI does not consider minority interests. Brandtzæg points out that the world has never before faced such an intrusive technology, which necessitates regulation and balancing against real human needs and values.
“We must not forget that AI is not a public, democratic project. It’s commercial, and behind it are a few American companies and billionaires,” says Brandtzæg.
More information:
Marita Skjuve et al, Unge og helseinformasjon, Tidsskrift for velferdsforskning (2025). DOI: 10.18261/tfv.27.4.2
Citation:
Media professor says AI’s superior ability to formulate thoughts for us weakens our ability to think critically (2025, November 16)
retrieved 16 November 2025
from https://techxplore.com/news/2025-11-media-professor-ai-superior-ability.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no
part may be reproduced without the written permission. The content is provided for information purposes only.
As nearly half of all Australians say they have recently used artificial intelligence (AI) tools, knowing when and how they’re being used is becoming more important.
Consultancy firm Deloitte recently partially refunded the Australian government after a report they published had AI-generated errors in it.
Amid these examples, a range of “AI detection” tools have emerged to try to address people’s need for identifying accurate, trustworthy and verified content.
But how do these tools actually work? And are they effective at spotting AI-generated material?
How do AI detectors work?
Several approaches exist, and their effectiveness can depend on which types of content are involved.
Detectors for text often try to infer AI involvement by looking for “signature” patterns in sentence structure, writing style, and the predictability of certain words or phrases being used. For example, the use of “delves” and “showcasing” has skyrocketed since AI writing tools became more available.
However the difference between AI and human patterns is getting smaller and smaller. This means signature-based tools can be highly unreliable.
Detectors for images sometimes work by analyzing embedded metadata which some AI tools add to the image file.
For example, the Content Credentials inspect tool allows people to view how a user has edited a piece of content, provided it was created and edited with compatible software. Like text, images can also be compared against verified datasets of AI-generated content (such as deepfakes).
Finally, some AI developers have started adding watermarks to the outputs of their AI systems. These are hidden patterns in any kind of content which are imperceptible to humans but can be detected by the AI developer. None of the large developers have shared their detection tools with the public yet, though.
Each of these methods has its drawbacks and limitations.
How effective are AI detectors?
The effectiveness of AI detectors can depend on several factors. These include which tools were used to make the content and whether the content was edited or modified after generation.
The tools’ training data can also affect results.
For example, key datasets used to detect AI-generated pictures do not have enough full-body pictures of people or images from people of certain cultures. This means successful detection is already limited in many ways.
Watermark-based detection can be quite good at detecting content made by AI tools from the same company. For example, if you use one of Google’s AI models such as Imagen, Google’s SynthID watermark tool claims to be able to spot the resulting outputs.
But SynthID is not publicly available yet. It also doesn’t work if, for example, you generate content using ChatGPT, which isn’t made by Google. Interoperability across AI developers is a major issue.
AI detectors can also be fooled when the output is edited. For example, if you use a voice cloning app and then add noise or reduce the quality (by making it smaller), this can trip up voice AI detectors. The same is true with AI image detectors.
Explainability is another major issue. Many AI detectors will give the user a “confidence estimate” of how certain it is that something is AI-generated. But they usually don’t explain their reasoning or why they think something is AI-generated.
It is important to realize that it is still early days for AI detection, especially when it comes to automatic detection.
A good example of this can be seen in recent attempts to detect deepfakes. The winner of Meta’s Deepfake Detection Challenge identified four out of five deepfakes. However, the model was trained on the same data it was tested on—a bit like having seen the answers before it took the quiz.
When tested against new content, the model’s success rate dropped. It only correctly identified three out of five deepfakes in the new dataset.
All this means AI detectors can and do get things wrong. They can result in false positives (claiming something is AI generated when it’s not) and false negatives (claiming something is human-generated when it’s not).
For the users involved, these mistakes can be devastating—such as a student whose essay is dismissed as AI-generated when they wrote it themselves, or someone who mistakenly believes an AI-written email came from a real human.
It’s an arms race as new technologies are developed or refined, and detectors are struggling to keep up.
Where to from here?
Relying on a single tool is problematic and risky. It’s generally safer and better to use a variety of methods to assess the authenticity of a piece of content.
You can do so by cross-referencing sources and double-checking facts in written content. Or for visual content, you might compare suspect images to other images purported to be taken during the same time or place. You might also ask for additional evidence or explanation if something looks or sounds dodgy.
But ultimately, trusted relationships with individuals and institutions will remain one of the most important factors when detection tools fall short or other options aren’t available.
Citation:
How do ‘AI detection’ tools actually work? And are they effective? (2025, November 16)
retrieved 16 November 2025
from https://techxplore.com/news/2025-11-ai-tools-effective.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no
part may be reproduced without the written permission. The content is provided for information purposes only.
Tech7 days ago
Tech2 days ago
Politics3 days ago
Sports2 days ago
Sports1 week ago