Guide: AI in business › PR and communication in the age of AI
Sztuczna inteligencja zmieniła komunikację kryzysową na dwa sposoby: przyspieszyła start kryzysów wizerunkowych i wydłużyła ich koniec. Fałszywe nagranie prezesa albo zmyślony komunikat można dziś zrobić w kwadrans. A gdy kryzys w mediach wygaśnie, modele AI jeszcze długo mogą powtarzać jego wersję w odpowiedziach o firmie. Opisujemy nowe rodzaje zagrożeń i skalę zjawiska w Polsce, pokazujemy, w czym AI pomaga zespołowi kryzysowemu, a czego nie wolno jej oddać, i dajemy plan na pierwszą godzinę.
In 60 seconds An AI crisis in five sentences
- What has changed: crisis communication is still a company’s plan and way of speaking at a moment of threat — except that today a crisis is increasingly started by a deepfake or false information.
- Who is targeted: companies with a recognisable board, a consumer brand or a large finance department. That is where impersonation turns into money fastest.
- What pays off: preparation before a crisis, i.e. monitoring, statement templates and drills. Improvisation is the most expensive option.
- First move: a list of people who can confirm or deny a recording within minutes, and an alert threshold in your monitoring.
- The limit for the machine: AI analyses and prepares, but no statement goes out without human approval.
What is crisis communication when AI is in play?
Crisis communication is the planned way in which a company informs people when its reputation, customer safety or business continuity are at risk. An image crisis begins when more people are talking about an issue than the company can handle in its normal way of working. We have gathered the classic principles in our guide on crisis PR management, and we have devoted a separate article to recognising the first signals: how to recognise and contain an image crisis.
AI doesn’t invalidate any of this. It changes the pace and the evidence. A crisis used to start with a real event or a real document; today it can start with a recording nobody ever filmed and a statement nobody in the company wrote. How this topic fits into the wider use of AI in an organisation is explained in the guide with which we open the topic of AI in business.
What crises has AI brought?
Four types of threat we see most often:
Każdy z tych rodzajów zostawia inne ślady, więc plan komunikacji kryzysowej powinien mieć osobny scenariusz dla każdego, a nie jeden wspólny na „atak w sieci”.
- Deepfake. A realistic, artificially generated recording or photo, for example of a CEO announcing redundancies or a product that ‘explodes’. We described how to recognise them in our article on deepfakes and how to detect them.
- Fake statements and documents. A forged press release, a letter bearing the ‘company logo’, a screenshot of a non-existent email.
- Mass fake reviews and comments generated automatically to mimic grassroots outrage.
- The memory of AI models. ChatGPT, Gemini or Perplexity may cite old allegations for months if there is a lack of up-to-date, credible sources with an explanation online.
How big is the problem of deepfakes and disinformation in Poland?
These are no longer isolated incidents. According to NASK data cited by Telix.pl, in May 2026 13.2k ads with deepfakes were identified, using the likenesses of 380 well-known people. In a survey by the Quantify agency of 1,000 Polish adults, 66% failed to recognise a fake AI-generated voice recording. Two thirds.
The most high-profile Polish case is the dispute between Rafał Brzoska and Omenaa Mensah and Meta over fake ads using their likenesses. According to PRoto.pl, on 29 April 2026 the Court of Appeal in Warsaw held the platform liable for such ads, rejecting the argument that it is merely a neutral intermediary.
For finance departments, the most instructive case is that of Arup. As CNN reported in May 2024, an employee at the Hong Kong office transferred around 25 million dollars in 15 transactions after a video call in which the CFO and the other participants were generated by AI. None of the company’s systems was hacked. What failed was the procedure for confirming instructions.
Which companies are most at risk?
The risk is not evenly spread. The most frequent targets are organisations in which a fake quickly turns into money or into customer fear:
Jeśli rozpoznajesz swoją firmę w dwóch punktach z tej listy, deepfake prezesa albo fałszywy komunikat przestaje być abstrakcją, a staje się pozycją w rejestrze ryzyk z właścicielem i terminem przeglądu.
- companies with a recognisable CEO or owner whose likeness is easy to fake,
- consumer brands (food, cosmetics, toys, automotive), where a rumour about a hazard has an immediate effect,
- companies with a large finance department and many branches, where a transfer instruction passes through several people,
- listed companies and companies in the middle of a transaction, where a fake statement can move the share price or negotiations,
- employers undergoing restructuring, when a recording about redundancies falls on fertile ground.
Jak Commwork®AI skraca drogę od fałszywki do reakcji
W kryzysie z udziałem AI wygrywa ten, kto pierwszy wie i pierwszy mówi. Commwork®AI łączy monitoring z ośmioma warstwami analizy, ocenia powagę sygnału i przygotowuje szkic oświadczenia, a zespół kryzysowy Commplace® prowadzi komunikację. Każde oświadczenie zatwierdza człowiek.
- social media 44%
- Forums and comments 21%
- Komunikatory i poczta 17%
- National media 11%
- Responses from AI models 7%
- 2.08.2026Art. 50 AI Act: obowiązek oznaczania deepfake’ów
- Week 2Mapa zagrożeń z czterema scenariuszami gotowa
- Week 4Monitoring nazwisk zarządu z progiem alarmowym działa
- Tydzień 5Wzory oświadczeń wstępnych czekają na prawnika
- Tydzień 10Ćwiczenie z fałszywym nagraniem na czas
- Week 12Aktualizacja księgi kryzysowej po ćwiczeniu
- 01Threat map
- 02Progi alarmowe
- 03Wzory oświadczeń
- 04Ćwiczenie z fałszywką
- 05Pomiar po kryzysie
- Po monitoringu dokłada osiem warstw: wagę źródła, dotarcie do Twoich grup, lukę wobec konkurencji i prognozę, czy temat urośnie.
- Przy nagłym wzroście negatywnych wzmianek ocenia powagę, zwołuje zespół i przygotowuje oświadczenie do podpisu; nic nie wychodzi bez akceptacji człowieka.
- Sprawdza w ośmiu silnikach AI, czy odpowiedzi o firmie nadal powtarzają zarzuty, a sposób wyboru narzędzia nasłuchu opisaliśmy w tekście o AI media monitoring.
The view reproduced is modeled after the dashboards provided to customers. The names, numbers, and terms are examples and show the way the system works, not the results of a specific company.
How does AI help a crisis team?
AI is a good helper in the first minutes because it reads a lot quickly. It is less good at judging what this means for people and for the company. Hence the division of labour:
| Task | What AI does | What stays with the human |
|---|---|---|
| Early detection | monitors media and the web around the clock, catches sudden spikes in mentions | judging whether it's a crisis or just noise |
| Tone assessment | groups mentions by tone and topic (sentiment analysis) | checking a sample, because AI confuses irony with praise |
| Drafts | prepares statement variants and answers to questions | selection, editing, approval by the lawyer and the board |
| Scenarios | simulates how media and customers might react | deciding which scenario is realistic |
| Verifying material | helps identify signs of manipulation in a recording or photo | confirmation at the source: did this person actually say it |
How to respond in the first hour, step by step
The first hour decides whose version of events spreads. The plan below works both for a deepfake and for a real incident.
- STEP 01Minutes 0–15: assess the scale
Who is talking, where, how fast it is growing, whether it concerns people’s safety. Assign a threat level according to the table below.
- STEP 02Minutes 15–30: verification
Czy materiał jest prawdziwy? Przy nagraniu potwierdź u osoby, którą przedstawia, i ustal pierwsze miejsce publikacji. Zabezpiecz dowody: zrzuty, adresy, godziny. Kto co robi w takiej sytuacji, najlepiej ustalić wcześniej, w księdze kryzysowej.
- STEP 03Minutes 30–45: a short holding statement
What we know, what we are checking, when we will say more. No speculation. AI can prepare versions; a human chooses and approves.
- STEP 04Minutes 45–60: channels
The company website as the single permanent place for updates, plus social media, journalists, employees and partners. For a fake, a report to the platform on which it appeared.
- STEP 05After the first hour: a rhythm of updates
Fixed times for subsequent statements and a decision log: who approved what and when. After the crisis, this log will prove to be the most valuable material for lessons learned, because it will show where you really lost time.
| Level | Example | Who decides |
|---|---|---|
| 1. Noise | a few negative comments with no reach | communications team, monitoring |
| 2. Problem | an article in trade media, a growing discussion | head of communications |
| 3. Crisis | national media, a deepfake involving the board | crisis team with the board and a lawyer |
| 4. Threat | people's safety, an investigation, a data leak | board, lawyer, emergency services; communication subordinate to safety |
What should you do when ChatGPT repeats false information about your company?
Stare zarzuty wracają w odpowiedziach AI tym częściej, im dłużej nikt ich nie prostował w sieci. Jak przerwać ten mechanizm u źródła, pokazujemy w tekście o tym, jak zakończyć spiralę fałszywych informacji na temat firmy.
Punktem wyjścia jest pomiar: ta sama lista kilkunastu pytań zadawana co miesiąc, tak jak w AI visibility audits. Bez niego nie wiesz, czy ChatGPT nieprawdziwe informacje o firmie nadal powtarza, czy już przestał.
An AI model repeats what it finds in sources or has memorised from training data. There is simply no ‘correct the answer about my company’ form that guarantees a change. What works is working at the source:
- publish a clear, dated explanation with facts on your own website, written so that it can be quoted in a single sentence,
- make sure there are corrections and up-to-date publications in credible media, because models trust independent sources more often (we describe the mechanism in our article on AI PR and visibility in ChatGPT),
- report an incorrect answer using the rating button in the given tool; it is a signal to the provider, not a guarantee,
- measure every month, on the same list of questions, whether the answers are changing.
How much does it cost to prepare a company for an AI-driven crisis?
The cost of preparation can be predicted. The cost of improvisation cannot. The table shows what the budget consists of and what drives it up; we don’t give rates, because they depend on company size and the number of markets, but the items themselves look similar in every organisation.
Największa pozycja w tym budżecie rzadko jest widoczna na fakturze. To czas zarządu, który bez planu spędza pierwszą godzinę kryzysu na ustalaniu, kto ma mówić. Jeśli chcesz zacząć od mapy zagrożeń i wzorów na pierwszą godzinę, zobacz, jak wygląda przygotowanie firmy w ramach zarządzania kryzysowego, a gdy potrzebujesz przede wszystkim dokumentu, sprawdź zakres opracowania księgi kryzysowej.
| Item | What it covers | What drives the cost |
|---|---|---|
| Monitoring with alerts | media, web, board members’ names, alert threshold | number of sources, keywords and languages |
| Threat map | list of scenarios, including deepfakes and fake documents | complexity of operations and number of brands |
| Crisis manual | roles, channels, statement templates, contacts | number of companies, countries and approvers |
| Drill | simulation with a fake recording and time pressure | number of participants and realism of the scenario |
| Crisis support | team available outside working hours | on-call arrangements and response time |
When to prepare and how often to practise?
The best moment is before the first signal. Straight away, set up round-the-clock monitoring with an alert threshold, ready-made holding statement templates and a list of people who can confirm or deny material within minutes. For finance, add one rule: no transfers or decisions based on voice or video alone, always a second confirmation channel.
Próg alarmowy, dyżur i wzory oświadczeń działają najlepiej, gdy leżą w jednym miejscu. W kokpicie Commwork®AI nagły wzrost negatywnych wzmianek uruchamia ocenę powagi, zwołanie zespołu i szkic oświadczenia, który czeka na podpis człowieka. Fałszywka zwykle rozlewa się najpierw w mediach społecznościowych, więc przejrzyj też nasze zasady na kryzysy w social media.
Once a quarter, run through a deepfake scenario. More often ahead of an AGM, a launch, a restructuring, elections or a major industry event, because that is when fakes have the greatest impact. We have gathered examples of solutions from Polish companies in our article crisis situations in practice. We write more broadly about the risks AI brings in our article on the risks and ethics of AI in business.
Gotowość na kryzys w 90 dni: lista kontrolna i próg decyzji
Scenariusz ilustracyjny, bez prawdziwej firmy w tle. Weźmy producenta z rozpoznawalnym prezesem, trzema oddziałami i działem finansów, który wysyła kilkadziesiąt przelewów tygodniowo. W pierwszym miesiącu zespół spisuje scenariusze, w tym deepfake z udziałem zarządu i podrobiony komunikat, ustawia monitoring nazwisk i marek z progiem alarmowym oraz listę osób, które potwierdzają materiały. W drugim przygotowuje oświadczenie kryzysowe w trzech wersjach (dementi, „sprawdzamy”, przeprosiny) i zasadę drugiego kanału dla przelewów. W trzecim ćwiczy scenariusz z fałszywym nagraniem na czas, a wnioski wpisuje do księgi. Efekt po 90 dniach da się sprawdzić: czas od sygnału do pierwszego oświadczenia zmierzony na ćwiczeniu, a nie deklarowany.
Zanim uznasz, że firma jest gotowa, przejdź przez listę. Każdy punkt ma mieć nazwisko osoby i datę ostatniego sprawdzenia.
Kiedy sięgnąć po pomoc z zewnątrz? Jeśli po odhaczeniu listy brakuje Ci więcej niż dwóch punktów albo nikt w firmie nie ćwiczył reakcji na fałszywe informacje o firmie od roku, pierwsza godzina kryzysu będzie improwizacją. Wtedy lepiej ściągnąć zewnętrzne wsparcie teraz niż szukać go w poniedziałek o 7:40. Wybierając agencję albo dostawcę monitoringu, zapytaj: kto odbierze telefon w nocy, ile minut mija od alarmu do pierwszego szkicu, jak weryfikujecie nagranie, kto podpisuje oświadczenie i czy mierzycie, co modele AI mówią o firmie po kryzysie. Dobre wzorce odpowiedzi zebraliśmy w tekście o tym, jak wygląda effective communication in a crisis situation, a przygotowanie z wyprzedzeniem opisujemy w poradniku, jak zabezpieczyć się przed kryzysem i wyjść z niego obronną ręką. Dezinformacja w Polsce rośnie szybciej niż budżety na jej odpieranie, więc ten przegląd rób co pół roku.
- monitoring nazwisk zarządu i marek z progiem alarmowym i osobą, która go odbiera,
- lista osób, które w kilka minut potwierdzą lub zdementują nagranie, z numerami poza firmową pocztą,
- trzy wzory oświadczenia wstępnego zatwierdzone przez prawnika,
- zasada drugiego kanału przy każdym poleceniu przelewu wydanym głosem lub wideo,
- stała strona z aktualizacjami, którą uruchomisz w kwadrans,
- zapisany wynik ostatniego ćwiczenia i lista pytań, na które co miesiąc sprawdzasz odpowiedzi modeli AI.
Does AI content need to be labelled in crisis communications?
Article 50 of the AI Act has applied since 2 August 2026. A company that publishes a deepfake must disclose that the material has been artificially generated. AI-generated text on matters of public interest also requires labelling, unless it has undergone human editorial review and the company takes responsibility for it. A statement prepared with the help of AI, but checked and signed by a spokesperson, therefore usually does not require labelling.
Beware of a trap. An ‘illustrative’ graphic or video generated in a hurry for a denial may itself be deemed a deepfake and undermine the credibility of the explanation, so in a crisis use real photos and recordings. You will find details of the obligation in our piece on labelling AI content from 2 August 2026, and the rules for writing the statement itself match what we advise for press releases in the AI era. If you want to prepare your company with us, see how we run crisis management at Commplace®. For international clients: crisis communications agency in Poland.
What not to do when a crisis is fuelled by AI
A fake recording with emotional content gathers thousands of views in an hour. A company that holds back its comment until the explanations are complete hands the narrative to the authors of the fake and then has to fight an entrenched version.
AI text may contain false assumptions or wording that a lawyer would have stopped. One ill-considered sentence in a crisis can be quoted for years.
The Arup case showed that a video call is not proof of identity. The lack of a second-channel rule can cost millions, and once the loss comes to light, an image crisis follows on top.
Removing criticism without a response looks like a cover-up. Screenshots of deleted posts come back as evidence, and the crisis gets a second topic: censorship.
AI models may repeat the old version for months. Without up-to-date, credible sources with an explanation, a customer will ask ChatGPT about the company and get allegations from a year ago.
Questions boards ask us about fake recordings and crises
How can you recognise a deepfake involving your CEO?
Only confirmation at the source gives certainty: was this person in this place and did they say this. Warning signs include unnatural mouth movements, odd lighting on the face, an overly smooth voice and no original source of publication. Agree in advance a channel through which the board can confirm or deny recordings within minutes.
Does AI content need to be labelled in crisis communications?
Yes, if the company publishes a deepfake or AI-generated text on matters of public interest without editorial review. This follows from Article 50 of the AI Act, applicable since 2 August 2026. A statement prepared with the help of AI, but checked and signed by a spokesperson, usually does not require labelling.
How do you prepare a company for an AI-related crisis?
Before a crisis, you need round-the-clock monitoring with an alert threshold, ready-made holding statement templates and a list of people who can confirm or deny material within minutes. On top of that, a drill with a deepfake scenario once a quarter.
What is an image crisis?
An image crisis is a situation in which an event or information about a company starts damaging its reputation faster than the company can explain it in its normal way of working. It may start with a real mistake, but increasingly also with a fake recording or document.
How big is the scale of deepfakes in Poland?
According to NASK data described in May 2026, 13.2k ads with deepfakes using the likenesses of 380 well-known people were identified. In the Quantify survey, 66% of Polish adults failed to recognise a fake AI-generated voice recording.
What should you do when ChatGPT repeats old allegations against your company?
Publish a clear, dated explanation with facts on your website and make sure there are up-to-date publications in independent media, because models use sources like these. Report the incorrect answer with the rating button and check the same questions every month. There is no guarantee of a quick change.
Can AI run crisis communication for a company?
No. AI speeds up monitoring, groups mentions and prepares drafts, but it does not assess the consequences for people and does not take responsibility. Statements in a crisis are approved by a human, and in risky matters by two people.
Sources
- SOURCEEuropean Commission: transparency obligations under Article 50 of the AI Act (FAQ) · as of 24 September 2026
- SOURCE
- SOURCE
- SOURCE
Read next
Dostaniesz mapę zagrożeń z deepfake’ami, wzory oświadczeń na pierwszą godzinę i ćwiczenie z presją czasu. Zaczynamy od rozmowy o scenariuszach, które realnie grożą Twojej branży.
Sebastian Kopiej, President of Commplace®. In public relations since 1996. Text prepared with the help of AI tools and fact-checked by the author. Data current as of September 29, 2026.