The launch of ChatGPT in late 2022 propelled Generative AI to prominence, gaining significant visibility and popularity among both consumers and organizations. Telecom operators have been quick to experiment with this technology, exploring applications that would boost individual productivity and optimize industry-specific processes, like client interactions

Most recently we have seen a significant interest in applying Generative AI to Network Operations, which is an area in which CSPs (Communications Service Providers) have long sought efficiency gains to enable the management of an increasingly complex technology landscape. In this article, we explore how operators can leverage Gen AI to augment the employee experience and automate business processes, all in the name of smarter, faster, more resilient networks.

Network knowledge at your fingertips

Generative AI is a gamechanger when it comes to intelligent document querying and information retrieval. The retrieval augmented generation (RAG) pattern allows the interrogation of text indexes via a semantic representation of a query and has quickly become a standard. Based on retrieved sections of selected documents, Generative AI can easily produce summaries compliant with pre-defined templates.

This application is demonstrated with one of our clients, who is now providing network technicians with a tool that enables them to quickly access a summary of past incidents that have affected a specific node on the broadband access network. Thanks to Generative AI, the new tool can retrieve all incidents relevant to the current investigation and produce a formatted summary of past events – reducing the time for this task.

CSPs are also investigating the use of Generative AI for managing contracts, such as interconnect and roaming contracts, or cell tower lease agreements. Through the help of Gen AI-enabled tools, users can quickly search for specific clauses or ask direct questions, such as “What are the security procedures for accessing this site?” or “What is the average price for site rentals in Madrid?” This removes the need for extensive, complex searches for the most up-to-date and relevant contracts and amendments. It also reduces the risk of error, especially when navigating large, complex document repositories that can span several years and many geographies.

The rise of conversational interfaces

For the past 60 years, most traditional applications have been using either command-line or graphical, menu-based interfaces. While regular users have mastered these tools, occasional users struggle. This is why some customers still prefer calling the contact center instead of using a mobile app!

Generative AI gives a new dimension to user interfaces, moving from predefined, rigid dialogs, to more free-flowing and intuitive conversations. This is relevant for some of the interfaces used when operating networks, where the user experience can be much improved.

For instance, we are currently defining a proof-of-concept with one of our CSP clients to support field technicians who perform interventions at customer homes and/or network points of interest. These technicians frequently need specialized help from team leads or colleagues while in the field; if this help cannot be provided on demand, then the service agent may need to schedule a follow up intervention.

Generative AI allows the development of a conversational bot that enables the technician to get the information they need about the specific location/services and technologies that will enable them to successfully troubleshoot or deploy services. A voice bot or a chat bot makes the job of the technician easier and quicker, allowing for faster issue resolution and avoiding repeat and costly truck rolls.

These service tools can also be combined with augmented reality (AR) applications, such as using a mobile phone to scan physical devices and generate relevant information. A variety of new interfaces that enable this type of support is made possible by the emergence of multimodal models such as Google’s Gemini, OpenAI’s GPT-4 and Mistral AI’s Mistral 7B.

An additional application focuses on network configuration. Specific intents, like boosting radio capacity at a stadium ahead of a major event, are set up through dedicated interfaces that require expertise on network management applications. By employing a conversational interface, network engineers can effortlessly grasp available capacity and make configurations for upcoming activities through friendly conversations with an agent. This conversational agent doubles as an advisor, drawing insights from historical activities and the present network status.

Autonomous Networks monitoring with a human touch

Moving to a higher level of network automation is critical for network operators. In fact, this is the key to improving service quality within a more complex technological environment without adding additional staff. AI is key to moving from human-managed networks, supported by insights from data, to AI-managed networks. Generative AI can thus complement other AI models, such as anomaly detection and classification.

While the industry has settled on the term “Autonomous Networks,” the goal is not complete autonomy. Configuration of intent is essential, and ongoing network monitoring for compliance is crucial. Even with a significant level of automation, human oversight remains imperative to ensure safety and maintain the quality of service.

Generative AI can produce a human readable summary of the status and activity in the network, allowing human agents to understand if and how the intent is satisfied. Even when operating networks with a high level of automation, human agents must be able to investigate, ask specific questions and get replies.

In the same way, an Autonomous Network’s reaction to an alarm or an anomaly must be defined in advance by a network engineer. On older generations of solutions, scripts must be developed and tested, which requires strong expertise. With Generative AI, natural language could be used to define responses to alarms, going as far as to extract appropriate remediation procedures from process documents. This approach allows human experts to review and make necessary adjustments.

Leading the way in an AI revolution

Like traditional AI, there are many use cases for Generative AI in the Telecom industry. In addition to individual agent productivity tools, Generative AI can be used to refine and streamline operational processes to better manage networks.

At Capgemini, we are now experimenting with leading CSPs on how to augment or automate existing workflows through AI technologies, helping them create a smarter, faster, more resilient network.

TelcoInsights is a series of posts about the latest trends and opportunities in the telecommunications industry – powered by a community of global industry experts and thought leaders.

As organizations begin seriously considering their post-pandemic futures, they now face the challenge of walking the tight rope between meeting growth objectives and building sustainable businesses. 

Over the last year, promises of long-term sustainability agendas have more than tripled, with pledges of zero carbon and carbon neutrality abounding. While many organizations are turning to technology to meet these targets and solve environmental issues, enterprises need to ensure their IT does not become a part of the problem.  

The era of sustainable tech is on the rise. Companies are now leveraging innovative, data-driven technology to simultaneously streamline operations, cut carbon emissions, and reduce their carbon footprint. Often perceived as a savior rather than a sinner, the production, use, and disposal of technology has an often-overlooked carbon footprint: an estimated 57.4 million tons of e-waste were generated worldwide in 2021 alone. The total is growing by an average of 2 Mt a year. 

Accelerated by the fiercely dynamic and competitive markets, more organizations are embracing digital transformation across their business. As a result, demand for computing power and data storage is on the rise and so too is the energy required to produce and run them. Curating sustainable enterprise-wide digital systems will be crucial for any business trying to balance the objectives of sustainable growth post-pandemic.   

But with so little awareness on the matter, the road to successful and sustainable IT needs a clear and rigorous roadmap. Our research has found crucial factors to consider when building and implementing sustainable IT strategies – let’s have a look through each one.  

•  Understanding the task  

When it comes to strategy, half of firms have defined an enterprise-wide sustainability approach, yet less than one in five (18%) have a comprehensive, sustainable IT strategy with well-defined goals and target timelines.  

Before a clear and robust framework can be rolled out, organizations need to get clued in on what they are dealing with. Our research revealed an alarming gap in awareness regarding the overall environmental impact of IT, with fewer than half of executives (43%) globally aware of their organization’s IT footprint. Many are confused about the true impact of IT. Only 34%, for example, know that the production of mobiles and laptops has a higher carbon footprint than the usage of these devices over their lifetime.  

Getting a clear understanding of the issue is the first critical stage for firms looking to develop sustainable strategies. Once the baselines and benchmarks of an enterprise’s environmental footprint have been marked out, organizations can then look to establish, implement, and monitor key performance indicators, targets, and frameworks. 

•  Engaged and informed employees  

Employees and leaders who are engaged with sustainability agendas drive greater progress. Even the most thorough strategies can come undone when those involved are not committed to the cause. Taking things one step further by developing a specialist sustainable IT team can provide streamlined purpose and coherence. Organizations must adopt the same mindset with employees as with consumers. People want to buy from companies with sustainable products and services. Likewise, employees want to work for such organizations. People are a critical component of sustainability transformation. Therefore, you must foster a culture that celebrates and promotes environmentalism, while trusting and empowering your people to contribute their own ideas. Those who have made sustainability a pillar of the organizational culture have seen greater progress.
Our research found that 60% of organizations are adopting sustainability to align with the demands of potential employees.

•  Sustainable software architecture 

Sustainability needs to be at the very center of an organization’s business. While emissions and output need to be carefully scrutinized, developing a sustainable software architecture is an imperative.  

Understanding the environmental consequences of software deployment and making decisions based on the carbon cost of infrastructure will ingrain sustainability into the foundations of an enterprise. Once the architecture is available, specific software modules within the structural design must be viewed from a sustainability perspective. For instance, organizations should empower their developers to understand the carbon cost of their software modules and use green coding to produce algorithms that have minimal energy consumption, at all times.

Upskilling developers in circular design will help product and design teams lessen their waste and thus their environmental footprints.

Sustainable IT can play a critical role in creating a circular economy by reducing waste, maximizing resource efficiency, and promoting more sustainable production and consumption practices. By introducing sustainability into the company’s value chain, you will drive the whole organization toward new efficiencies and a circular economy.

Moving forward, sustainability must be at the core of all our efforts. While many organizations have begun to focus on their overall sustainability agenda, the critical issue of sustainable IT has been overlooked. To give sustainable IT the attention it deserves, organizations need to understand the carbon cost of our digital world and accelerate the move to sustainable systems with engaged and dedicated teams. In this way, sustainable IT can play a central part in tackling climate change, promoting a circular economy, driving innovation, and moving the world to a more resilient and sustainable future.  

The hype in AI is about whether a truly intelligent AI is an existential risk to society. Are we heading for Skynet or The Culture? What will the future bring?

I’d argue that the larger and more realistic threat is from Deep Stupidity — the weaponization of Artificial General Intelligence to amplify misinformation and create distrust in society.

Social media is the platform, AI is the weapon

One of the depressing things about the internet is how its made conspiracy theories spread. Where before people were lone idiots, potentially subscribing to some bizarre magazine or conspiracy society in a given area, you really didn’t have the ability to industrial scale these things. Social media and the Internet has increased the spread of such ideas. So while some AI folks talk about the existential threat of AGI, personally I’m much more concerned about Artificial General Stupidity.

So I thought it is worth looking at why it is much easier to build an AI that is a flat earther than it is to build a High School physics teacher, let alone a Stephen Hawking.

It is easier being confidently wrong and not understanding

LLMs are confidently wrong, that inability to actually understand is a great advantage when being a conspiracy theorist. Because when you understand stuff, then conspiracy theories are dumb.

This means the training data set for our AI conspiracy theorist must be incomplete, what we need is not something that has access to a broad set of data, but actually something that has access to an incredibly small and specific set of data that repeats the same point over and over again.

To be a conspiracy theorist means denying evidence and ignoring contradictions, this is much easier to learn and code for than actually receiving new information that challenges your current model and altering it.

Small data set for a single topic

So this is a massive advantage for LLMs when trying to create a conspiracy theorist. What we need is a limited set of data that repeats a given conclusion and continually lines up all evidence to that conclusion. We can apply this to lots of conspiracy theorists out there, for instance those folks who scream “false flag” after every single mass shooting incident in the US, in other words we have a small set of data, possibly only a few hundred data points that always result in the same conclusion. This means for our custom trained conspiracy theorist the association it always knows is “what ever the data, the answer is the conspiracy”.

Now we could get fancy and have a number of conspiracies, but given very few of them are logically consistent with each other, let alone with reality, it is more effective to have a model per conspiracy and just switch between them. That a conspiracy theorist is inconsistent with what they’ve previously said isn’t a problem, but we don’t want inconsistencies between conspiracies on a single topic. What we need to add are the standard “rebuttals of reality” like “Water finds its level”, “We don’t see the curve”, “NASA is fake” or “Spurs are a top Premier League club”.

Hallucinations help

This small set of data really helps us take advantage of the largest flaw in LLMs, hallucinations, or when the LLM just makes stuff up either because it has no data on the topic, or because the actual answer is rare so the weightings bias it towards an invalid answer. This is where LLMs really can scale conspiracy theories, because the probabilities are weighted towards the conspiracy theory already (as that is the only “correct” answer within the model) then any information we are provided with is recast within that context. So if someone tells us that the Greeks proved the earth was round in the 2nd Century BC our LLM could easily reply:

Context makes hallucinations doubly annoying

Our LLMs can go beyond the average conspiracy theorist thanks to the context and hallucinations. While an average conspiracy person will only have a fixed set of talking points, and potentially be constrained at some level by reality, the hallucinations and context of the conversation enables our conspiracy LLM to keep building its conspiracy and adding elements to it. Because our LLM is unconstrained by reality and counter arguments, instead being able to reframe any counter argument by using a hallucination it will be significantly more maddening. It will also mean it will create new justifications for the conspiracy that have never been put forwards before. These will, of course, be total nonsense but new total nonsense is mana from heaven to other conspiracy theorists.

Reset and start again

The final piece that makes a conspiracy LLM much easier to create is that if the LLM goes truly bonkers and you need to reset… this is exactly what conspiracy theorists do today. So if our LLM is creating hallucinations that fail some form of basic test, or just every 20 responses, we can reset the conversation in a totally different direction. Making my generative LLM detect either a frustration or an “ah ha” moment from the person it is annoying, a trivial task, enables me to then have my conspiracy bot just jump to another topic, and to do so in a much smoother way than most conspiracy theorists do today.

This is a much smoother transition for a flat earth conspiracy than you’ll hear on TikTok or YouTube.

We have achieved AGS, that isn’t a good thing

I’ve argued that the current generation of AIs aren’t close to genuinely passing the Turing test, let alone more modern tests. Turing set the bar of intelligence as the CEO of a Fortune 50 company, and made it have awareness of what it didn’t know.

Some folks are concerned about a coming existential crisis where Artificial General Intelligence becomes a threat to humanity.

But for me that is assuming the current generation of technologies are not a threat, and that intelligence is a greater threat than weaponized stupidity. Many people in AI are in fact arguing that GPT passes the Turing test, not because it replicates an intelligent human, but because either it can pass a multiple choice or formulaic example, or because it can convince people they are speaking to a not very bright person.

We can today make an AI that is the equivalent of a conspiracy theorist, someone untethered to reality and disconnected from logic. This isn’t General Intelligence, but it is General Stupidity.

Deep fakes and deep stupidity

Where Deep Fakes can make us not trust sources, Deep Stupidity can amplify misinformation and constantly give it justification and explanation. Where Deep Fakes imitate a person or event, Deep Stupidity can imitate the response of the crowd to that event. Spinning up a million conspiracy theorists to amplify not just the Deep Fake but the creation of an alternative reality around it.

The internet and particularly social media has proven a fertile ground for human created stupidity and conspiracy theories. Entire political movements and groups have been created based on internet created nonsense. These have succeeded in gaining significant mindshare without having the capacity to really generate either convincing material or convincing narratives.

AIs today have the ability to change that.

Stupidity and misinformation are today’s existential threat

We need to stop talking about the challenge with AI being only when it becomes “intelligent”, because it is already sufficiently stupid to have massive negative consequences on society. It is madness to think that companies, and especially governments, aren’t looking at this technologies and how they can use them to achieve their ends, even if their ends are simply to sew chaos.

Stupidity is the foundation for worrying about intelligence

Worrying about an AI ‘waking up’ and threatening humanity is a philosophical problem, but addressing Artificial Stupidity would give us the framework to deal with that future challenge. Everything about controlling and managing AI in future can be mapped to controlling and avoiding AGS today.

When we talk about frameworks for Trusted AI and legislation on things like Ethical Data Sourcing these are elements that apply to General Stupidity just as much as to intelligence. So we should stop worrying simply about some amorphous future threat and instead start worrying about how we avoid, detect and control Artificial General Stupidity, because in doing that we lay the platform for controlling AI overall.

This article first appeared on Medium.

Disclaimer: This blog was NOT written by ChatGPT, but by a group of human data scientists: Shahryar Masoumi, Wouter Zirkzee, Almira Pillay, Sven Hendrikx and myself.

Whether we are ready for it or not, we are currently in the era of generative AI, with the explosion of generative models such as DALL-e, GPT-3, and, notably, ChatGPT, which racked up one million users in one day. Recently, on March 14th, 2023, OpenAI released GPT-4, which caused quite a stir and thousands of people lining up to try it.

Generative AI can be used as a powerful resource to aid us in the most complex tasks. But like with any powerful innovation, there are some important questions to be asked… Can we really trust these AI models? How do we know if the data used in model training is representative, unbiased, and copyright safe? Are the safety constraints implemented robust enough? And most importantly, will AI replace the human workforce?

These are tough questions that we need to keep in mind and address. In this blog, we will focus on generative AI models, their trustworthiness, and how we can mitigate the risks that come with using them in a business setting.

Before we lay out our trust issues, let’s take a step back and explain what this new generative AI era means. Generative models are deep learning models that create new data. Their predecessors are Chatbots, VAE, GANs, and transformer-based NLP models, they hold an architecture that can fantasize about and create new data points based on the original data that was used to train them — and today, we can do this all based on just a text prompt!

We can consider chatbots as the first generative models, but looking back we’ve come very far since then, with ChatGPT and DALL-e being easily accessible interfaces that everyone can use in their day-to-day. It is important to remember these are interfaces with generative pre-trained transformer (GPT) models under the hood.

The widespread accessibility of these two models has brought about a boom in the open-source community where we see more and more models being published, in the hopes of making the technology more user-friendly and enabling more robust implementations.

But let’s not get ahead of ourselves just yet — we will come back to this in our next blog. What’s that infamous Spiderman quote again?

With great power…

The generative AI era has so much potential in moving us closer to artificial general intelligence (AGI) because these models are trained on understanding language but can also perform on a wide variety of other tasks, that in some cases even exceed human capability. This makes them very powerful in many business applications.

Starting with the most common — text application, which is fueled by GPT and GAN models. Including everything from text generation to summarization and personalized content creation, these can be used in education, healthcare, marketing, and day-to-day life. The conversational application component of text application is used in chatbots and voice assistants.

Next, code-based applications are fueled by the same models, with GitHub’s Co-pilot as the most notable example. Here we can use generative AI to complete our code, review it, fix bugs, refactor, and write code comments and documentation.

On the topic of visual applications, we can use DALL-e, Stable Diffusion, and Midjourney. These models can be used to create new or improved visual material for marketing, education, and design. In the health sector, we can use these models for semantic translation, where semantic images are taken as input and a realistic visual output is generated. 3D shape generation with GANs is another interesting application in the video game industry. Finally, text-to-video editing with natural language is a novel and interesting application for the entertainment industry.

GANs and sequence-to-sequence automatic speech recognition (ASR) models (such as Whisper) are used in audio applications. Their text-to-speech application can be used in education and marketing. Speech-to-speech conversion and music generation have advantages for the entertainment and video game industry, such as game character voice generation.

Although powerful, such models also come with societal limitations and risks, which are crucial to address. For example, generative models are susceptible to unexplainable or faulty behavior, often because the data can have a variety of flaws, such as poor quality, bias, or just straight-up wrong information.

So, with great power indeed comes great responsibility… and a few trust issues

If we take a closer look at the risks regarding ethics and fairness in generative models, we can distinguish multiple categories of risk.

The first major risk is bias, which can occur in different settings. An example of bias is the use of stereotypes such as race, gender, or sexuality. This can lead to discrimination and unjust or oppressive answers generated from the model. Another form of bias is the model’s word choice. Its answers should be formulated without toxic or vulgar content, and slurs.

One example of a language model that learned a wrong bias is Tay, a Twitter bot developed by Microsoft in 2016. Tay was created to learn, by actively engaging with other Twitter users by answering, retweeting, or liking their posts. Through these interactions, the model swiftly learned wrong, racist, and unethical information, which it included in its own Twitter posts. This led to the shutdown of Tay, less than 24 hours after its initial release.

Large language models (LLMs) like ChatGPT generate the most relevant answer based on the constraints, but it is not always 100% correct and can contain false information. Currently, such models provide their answers written as confident statements, which can be misleading as they may not be correct. Such events where a model confidently makes inaccurate statements are also called hallucinations.

In 2023, Microsoft released a GPT-backed model to empower their Bing search engine with chat capabilities. However, there have already been multiple reports of undesirable behavior by this new service. It has threatened users with legal consequences or exposed their personal information. In another situation, it tried to convince a tech reporter he was not happily married and that he was in love with the chatbot (it also proclaimed their love for the reporter) and consequently should leave his wife (you see why we have trust issues now?!).

Generative models are trained on large corpora of data, which in many cases, is scraped from the internet. This data can contain private information, causing a privacy risk as it can unintentionally be learned and memorized by the model. This private data not only contain people, but also project documents, code bases, and works of art. When using medical models to diagnose a patient, it could also include private patient data. This also ties into copyright when this private memorized data is used in a generated output. For example, there have even been cases where image diffusion models have included slightly altered signatures or watermarks it has learned from their training set.

The public can also maliciously use generative models to harm/cheat others. This risk is linked with the other mentioned risks, except that it is intentional. Generative models can easily be used to create entirely new content with (purposefully) incorrect, private, or stolen information. Scarily, it doesn’t take much effort to flood the internet with maliciously generated content.

Building trust takes time…and tests

To mitigate these risks, we need to ensure the models are reliable and transparent through testing. Testing of AI models comes with some nuances when compared to testing of software, and they need to be addressed in an MLOps setting with data, model, and system tests.

These tests are captured in a test strategy at the very start of the project (problem formulation). In this early stage, it is important to capture key performance indicators (KPIs) to ensure a robust implementation. Next to that, assessing the impact of the model on the user and society is a crucial step in this phase. Based on the assessment, user subpopulation KPIs are collected and measured against, in addition to the performance KPIs.

An example of a subpopulation KPI is model accuracy on a specific user segment, which needs to be measured on data, model, and system levels. There are open-source packages that we can use to do this, like the AI Fairness 360 package.

Data testing can be used to address bias, privacy, and false information (consistency) trust issues. We make sure these are mitigated through exploratory data analysis (EDA), with assessments on bias, consistency, and toxicity of the data sources.

The data bias mitigation methods vary depending on the data used for training (images, text, audio, tabular), but they boil down to re-weighting the features of the minority group, oversampling the minority group, or under-sampling the majority group.

These changes need to be documented and reproducible, which is done with the help of data version control (DVC). DVC allows us to commit versions of data, parameters, and models in the same way “traditional” version control tools such as git do.

Model testing focuses on model performance metrics, which are assessed through training iterations with validated training data from previous tests. These need to be reproducible and saved with model versions. We can support this through open MLOPs packages like MLFlow.

Next, model robustness tests like metamorphic and adversarial tests should be implemented. These tests help assess if the model performs well on independent test scenarios. The usability of the model is assessed through user acceptance tests (UAT). Lags in the pipeline, false information, and interpretability of the prediction are measured on this level.

In terms of ChatGPT, a UAT could be constructed around assessing if the answer to the prompt is according to the user’s expectation. In addition, the explainability aspect is added — does the model provide sources used to generate the expected response?

System testing is extremely important to mitigate malicious use and false information risks. Malicious use needs to be assessed in the first phase and system tests are constructed based on that. Constraints in the model are then programmed.

OpenAI is aware of possible malicious uses of ChatGPT and have incorporated safety as part of their strategy. They have described how they try to mitigate some of these risks and limitations. In a system test, these constraints are validated on real-life scenarios, as opposed to controlled environments used in previous tests.

Let’s not forget about model and data drift. These are monitored, and retraining mechanisms can be set up to ensure the model stays relevant over time. Finally, the human-in-the-loop (HIL) method is also used to provide feedback to an online model.

ChatGPT and Bard (Google’s chatbot) have the possibility of human feedback through a thumbs up/down. Though simple, this feedback is used to effectively retrain and align the underlying models to users’ expectations, providing more relevant feedback in future iterations.

To trust or not to trust?

Just like the internet, truth and facts are not always given — and we’ve seen (and will continue to see) instances where ChatGPT and other generative AI models get it wrong. While it is a powerful tool, and we completely understand the hype, there will always be some risk. It should be standard practice to implement risk and quality control techniques to minimize the risks as much as possible. And we do see this happening in practice — OpenAI has been transparent about the limitations of their models, how they have tested them, and the governance that has been set up. Google also has responsible AI principles that they have abided by when developing Bard. As both organizations release new and improved models — they also advance their testing controls to continuously improve quality, safety, and user-friendliness.

Perhaps we can argue that using generative AI models like ChatGPT doesn’t necessarily leave us vulnerable to misinformation, but more familiar with how AI works and its limitations. Overall, the future of generative AI is bright and will continue to revolutionize the industry if we can trust it. And as we know, trust is an ongoing process…

In the next part of our Trustworthy Generative AI series, we will explore testing LLMs (bring your techie hat) and how quality LLM solutions lead to trust, which in turn, will increase adoption among businesses and the public.

This article first appeared on SogetiLabs blog.